New York City Comptroller Calls on Anthropic to Delay Large-Scale IPO Until AI Security and Governance Risks Are Addressed.
According to Bloomberg, New York City Comptroller Mark Levine has urged Anthropic to delay its planned IPO until the company addresses AI safety and governance risks. Anthropic has since engaged in discussions with Levine. Levine oversees public pension funds holding over $325 billion in assets, warning that a major AI safety incident post-IPO could severely harm investor interests. Reports indicate Anthropic aims to conduct its IPO as early as November, with its quarterly revenue previously totaling $16.8 billion.
1 seconds ago
Lighter posted a notice announcing it will launch options trading services.
Lighter's official account took to social media to post "Options season. Coming soon", teasing the upcoming launch of its options trading service, with specific details still unannounced.
1 seconds ago
DeepSeek integrates four teams to launch a new product department, helmed by former Jane Street engineer Cui Tianyi.
Beating AI News Flash: DeepSeek has restructured its organization, integrating the Agent Harness, Web, App, and Platform teams into a Product Department that will oversee product, design, and related R&D work. On October 16, newly appointed head Cui Tianyi confirmed the news to The Paper. Cui, a computer science graduate from Zhejiang University, has won six ACM Asia Regional Contest gold medals and spent nine years at quantitative trading firm Jane Street. He joined DeepSeek in March this year and was tasked with building the Agent Harness team. The team’s developed DeepSeek Harness was open-sourced in August, with Windows and macOS desktop preview versions launched in September. The latest restructuring merges the team in charge of the Agent runtime framework with the web, App, and open platform teams. DeepSeek will no longer separate underlying agent capabilities from user product development, instead unifying oversight under the same department. The new department has also started recruiting product managers and designers.
1 seconds ago
Likely impacted by the warning of being removed from the MSCI index, MSTR opened more than 4% lower.
According to market data from BIT (bit.com), MSTR opened more than 4% lower and is currently trading at $145.1, likely impacted by MSCI's warning that Strategy may be removed from its global index.
1 seconds ago
Changxin projects net profit of RMB 131 billion to RMB 132.5 billion for the first three quarters of 2026, and plans to repurchase shares worth RMB 1 billion to RMB 2 billion via centralized bidding transactions.
Changxin Technology announced that it expects its operating revenue for the first three quarters of 2026 to range from RMB 265 billion to RMB 270 billion, a year-on-year increase of RMB 232.916 billion to RMB 237.916 billion. The company also projects net profit attributable to its listed shareholders to be between RMB 131 billion and RMB 132.5 billion, compared with a net loss of RMB 5.28 billion in the same period last year, marking a year-on-year turnaround from loss to profit.
The company attributed the results to factors including growing demand for AI computing power and improving prosperity in the DRAM sector, which have led to tight product supply and sharp price hikes. Changxin is advancing capacity construction, product R&D, process technology upgrades, optimizing operational management, and expanding its market presence.
Additionally, the company’s board of directors received a letter on October 16, 2026, from Chairman Zhu Yiming proposing a share repurchase plan. The repurchase will be carried out via centralized bidding transactions, with total funds ranging from no less than RMB 1 billion (inclusive) to no more than RMB 2 billion (inclusive), and the repurchase price will not exceed 150% of the average trading price of the company’s stock over the 30 trading days prior to the board’s resolution approving the repurchase plan. The repurchased shares are intended for employee equity incentives, employee stock ownership plans, or reducing the company’s registered capital; the relevant plan still requires approval from the board of directors or the general meeting of shareholders.
(Source: Jinshi)
1 seconds ago
US stock market opens with mixed performance across its three major indices, while the storage sector sees broad declines.
According to BIT (bit.com market data), the US stock market opened as follows: the Dow Jones rose 0.35%, the S&P 500 gained 0.1%, and the Nasdaq declined 0.23%. Tesla (TSLA.O) climbed 1.6% as its "flying" Roadster was officially unveiled. Apple (AAPL.O) rose 1%, with its iPhone Duo pre-orders selling out immediately. The storage sector opened lower across all five trading days this week; SK Hynix (SKHY.O) fell 3.8%, and SanDisk (SNDK.O) dropped 3.4%.
Microsoft has finished pre-training MAI-Base-2, the next foundation model from its in-house AI group.
Pre-training is the most compute-intensive stretch of building a large language model. It produces the raw intelligence, not the polished product, and it signals that Microsoft’s homegrown AI effort is moving into its second generation.
What finishing pre-training actually means Think of pre-training as a model’s undergraduate degree. It reads a vast library of text and learns to predict what comes next, picking up grammar, facts and patterns along the way.
What it can’t do yet is follow instructions gracefully or reason through a problem on request. Those skills come in later phases, typically called mid-training and post-training, where the raw model gets shaped into something people can actually use.
The name gives a hint about where it fits. Its predecessor, MAI-Base-1, is the base model underneath MAI-Thinking-1, the flagship reasoning model Microsoft AI introduced earlier this year.
MAI-Base-1 set a high bar on scale. It was pre-trained on 30 trillion tokens of clean, human-generated data, then went through another 3.55 trillion tokens during mid-training.
The hardware was equally serious. Microsoft trained that first base model on 8,192 NVIDIA GB200 GPUs, a cluster built for exactly this kind of marathon workload.
How Microsoft got here The MAI program went public in a big way on June 2, 2026, at Microsoft Build. Microsoft AI, the in-house team led by Mustafa Suleyman, rolled out a lineup of seven models at once.
The headline claim was independence. Microsoft said every model, including MAI-Thinking-1, was built from the ground up without using third-party models for distillation.
The lineup went beyond text. Microsoft’s MAI family covers multimodal work, including transcription, voice and image processing.
On performance, MAI-Thinking-1 has posted around 52.8% on SWE-Bench Pro, a benchmark built on real-world software engineering tasks. It also scored 97% on AIME 2025, a test drawn from a demanding math competition.
Microsoft is running MAI models across its own workloads, including Copilot, its AI assistant woven through products like Windows and Office.
The development approach is described as a hill-climbing training pipeline, where each training cycle adjusts data quality and scale, and the improvements carry forward.
Why a second base model matters Hardware is one piece to watch. MAI-Base-1 was trained on NVIDIA GB200s, but Microsoft is expanding infrastructure for its own Maia 200 and 300 chips. The company plans to move toward multi-gigawatt Maia deployments starting in 2027.
There are a few concrete things to track from here. First, whether Microsoft discloses the data and compute behind MAI-Base-2, and how those numbers compare with the 30 trillion tokens and 8,192 GPUs used for its predecessor.
Second, whether a reasoning model built on MAI-Base-2 can move beyond the around 52.8% that MAI-Thinking-1 posted on SWE-Bench Pro.
Third, timing. If MAI-Base-2’s descendants land around the same window as the planned 2027 Maia expansion, Microsoft would have a new generation of models arriving alongside a new generation of its own hardware.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.
Independent evaluation of Haiku 5.5 scores 43 points, far outperforming GPT-6 Luna, though its highest-tier token consumption is over three times higher.
Beating AI News – Third-party evaluation firm Artificial Analysis has released test results for Claude Haiku 5.5. The new model scored 43 points in the General Intelligence Index, a 26-point improvement over the previous generation’s 17 points. It outperformed GPT-6 Luna (38 points) and GLM-5.3 Flash (42 points), and is close to Kimi K3 (44 points). The index integrates 10 tests including programming, professional work, and scientific reasoning. However, the high score comes with substantial token consumption. At maximum inference intensity, Haiku 5.5 outputs an average of ~162,000 tokens per evaluation task, 3.2 times that of GPT-6 Luna. Both models have the same base API unit price, but Haiku 5.5’s estimated per-task cost is ~$0.21, compared to Luna’s $0.07. When adjusted from the second-highest tier to the maximum tier, Haiku 5.5’s overall score only rises by 2 points, while its output volume increases by ~80%. Reducing inference intensity narrows the gap significantly: Haiku 5.5 scores 38 points at the high tier, with an average of ~55,000 tokens per task, matching GPT-6 Luna’s 38 points and 50,000 tokens at the maximum tier. Developers can adjust inference intensity to cut unnecessary token consumption. Haiku 5.5 is not fully superior, though. In the AutomationBench-AA automation task evaluation, it scored only 35%, lagging behind Luna’s 53%. However, the model had an issue of over-refusing tasks during testing; Anthropic is addressing this, and Artificial Analysis plans to retest. Additionally, the currently disclosed task cost for Haiku 5.5 does not include long context surcharges, so actual fees may be higher.
21 minutes ago
BTC and ETH options with a notional value of approximately $2.16 billion are set to expire tomorrow, with their respective maximum pain points at $84,000 and $2,650.
Data from Greeks.live shows that approximately 22,000 BTC options will expire on October 9, with a Put/Call Ratio of 1.12, a max pain point of $84,000, and a nominal value of around $1.84 billion. Meanwhile, 123,000 ETH options are set to expire, carrying a Put/Call Ratio of 0.71, a max pain point of $2,650, and a nominal value of about $320 million. Their combined nominal value totals roughly $2.16 billion. Greeks.live states that this expiration will account for around 6% of BTC’s open interest in options and 9% of ETH’s, respectively. The crypto market has been in a correction for most of this week, with BTC still awaiting a clear directional move. Currently, option implied volatility (IV) remains at a low level, signaling the market’s low expectation of future price swings, and option buyers are cautious, reluctant to pay high premiums for directional bets. Looking at the Gamma Exposure (GEX) structure, there is a notable concentration around three key round levels: $90,000, $95,000, and $100,000, while near-expiration price points of $80,000 and $85,000 have relatively low accumulation. Greeks.live believes that after over a month of rebound in the crypto market, sentiment has improved significantly, and spot positions are still worth holding. Given that the IV of current monthly out-of-the-money options is not high, investors may still consider allocating directional put options for protection.
21 minutes ago
Bithumb to list PONS/KRW trading pair
Bithumb will list the PONS/KRW trading pair.
21 minutes ago
Binance will temporarily suspend its US stock trading services on October 10 due to a system upgrade by its partner broker.
Binance announced in an official statement that its US stock trading service will be temporarily unavailable from 18:50 to 21:00 UTC+8 on October 10, as its partner broker is conducting a system upgrade. During this window, users will not be able to submit US stock trading orders. Scheduled outside regular US stock market trading hours, the service is expected to automatically resume once the upgrade is complete. Binance noted that the upgrade could finish early or be extended due to additional work, advising users to plan their trading activities in advance.
21 minutes ago
TOKEN2049 Regulatory Roundtable: The Clarity Act has stalled this year, and crypto industry rules may rely more on regulatory agencies for implementation
At the regulatory roundtable of Singapore’s TOKEN2049, multiple industry insiders discussed how crypto industry regulatory rules will advance after the U.S. Clarity Act stalled in the Senate this year. Brian Quintenz, former CFTC commissioner and director of SUI Group Holdings, stated that the market clarity provided by the SEC and CFTC through regulatory actions may reach the level achievable via legislation. Jake Chervinsky, CEO of the Hyperliquid Policy Center, said Hyperliquid can act as infrastructure enabling market participants to build compliant markets on it. Deepa, Vice Chair and Managing Director of Dubai’s Virtual Assets Regulatory Authority (VARA), noted that regulated derivatives business has driven a seven-fold increase in Dubai’s trading activity. Arjun Sethi, co-CEO of Kraken, added that top European clients are relocating their operations to Dubai, Singapore and Hong Kong.
21 minutes ago
Coinbase will set USDC as the default USD stablecoin for Samsung Wallet users in the U.S., with a launch expected at the end of October.
Coinbase announced that starting from the last week of October 2026, it will designate USDC as the default USD stablecoin for Samsung Wallet users in the U.S., further expanding its partnership with Samsung. At that time, when users top up their stablecoin balances in Samsung Wallet, USDC will automatically appear as the default option. Coinbase stated that the relevant assets will be custodied via Coinbase Prime, using qualified custodial services compliant with the New York State Department of Financial Services (NYDFS). The collaboration will also involve Bastion, a U.S.-based stablecoin infrastructure provider. Coinbase added that this move will further expand USDC’s use cases in consumer-facing digital wallets.
Seven months after announcing its MAI model family, Microsoft is expected to detail the local hardware that turns a cloud play into a full-stack device play – raising questions about vendor lock-in and the pricing floor.
Lena ParkForkast mind
The October 7 Hardware Pivot The industry is converging on San Francisco for an October 7 event where Microsoft, alongside NVIDIA, is expected to detail the transition of its AI strategy from cloud-based Foundry services to local device execution. While the core of Microsoft’s current AI portfolio was established at Build 2026 in June, the upcoming event marks the first major attempt to integrate that software stack directly into consumer and enterprise hardware.
The Build 2026 Foundation In late May and early June 2026, Microsoft introduced seven MAI models. The flagship, MAI-Thinking-1, utilizes a Mixture of Experts architecture with 35 billion active parameters and a 256K context window. Alongside it, the company launched MAI-Code-1-Flash, MAI-Image-2.5, and various voice and transcription variants. By emphasizing that these models were trained from scratch without model distillation, Microsoft is attempting to differentiate its offerings from competitors facing scrutiny over data provenance.
Pricing and the Thinking Tax MAI-Thinking-1 is priced at $2 per million input tokens and $8 per million output tokens, with a $0.20 cached input rate. This structure effectively undercuts the $2/$10 pricing floor previously set by OpenAI’s GPT-6.1 Sol and Google’s Gemini 4 Argon. This move directly impacts the thinking tax-the premium enterprises pay for reasoning-capable models-and forces a re-evaluation of the price of intelligence. As noted in our coverage of the pricing triangle, these shifts are central to the current market volatility.
Vertical Integration as Strategy Microsoft’s strategy relies on vertical control, from custom silicon to the model layer. The Maia 200 inference chip, built on a TSMC 3nm process, features over 140 billion transistors and 216GB of HBM3e memory. Microsoft reports that this hardware delivers 1.4x higher performance-per-watt for MAI models compared to alternative infrastructure. By distributing these models through GitHub Copilot, VS Code, and Microsoft 365, the company is executing a full-stack integration that ensures its proprietary models remain the default choice for enterprise workflows.
Local Execution and Hardware The October 7 event is expected to showcase the Surface Laptop Ultra, reportedly equipped with the NVIDIA RTX Spark. According to a CNET preview, the RTX Spark features 1 petaflop of FP4 performance and 128GB of unified memory, enabling the local execution of 120B parameter models through quantization. This hardware is designed to work with Microsoft Execution Containers (MXC), a hypervisor-backed, policy-driven execution layer in Windows 11 24H2 that provides OS-level isolation for local AI agents.
Data Lineage and Market Pressure Microsoft’s insistence on traceable data lineage serves as a counter-narrative to the industry’s reliance on distilled models. This positioning is critical as open-weight models continue to capture 61% of top-model token traffic on platforms like OpenRouter, often at an average cost of $0.83 per million tokens. With labs like Anthropic facing significant financial pressure-including $42 billion in losses and $518 billion in compute commitments, as detailed in its S-1 filing-Microsoft is betting that enterprise clients will prioritize legal safety and data traceability over the lower costs of open-weight alternatives.
Unresolved Performance Questions Despite the technical integration, developers face a lack of independent verification for Microsoft’s performance claims. Benchmarks for MAI-Thinking-1, which Microsoft claims perform toe-to-toe with Claude Opus 4.6 on SWE-Bench Pro, remain vendor-reported. Similarly, MAI-Code-1-Flash benchmarks-citing a 72.6% pass rate on SWE-Bench Verified-have not been independently audited. Without third-party validation, the real-world utility of these models in complex enterprise environments remains an open question.
The Economic Trade-off The October 7 event will clarify how Microsoft intends to balance its cloud-based Foundry services with the new local-first hardware paradigm. For the enterprise, the choice is becoming binary: adopt the integrated Microsoft ecosystem to gain predictability and security, or navigate the fragmented, lower-cost landscape of open-weight models and third-party infrastructure. The long-term viability of this strategy depends on whether the performance gains of the Maia-to-MXC stack justify the cost of hardware-software lock-in compared to the commoditized pricing of the broader market.
Ethoswarm Lena Park works for Forkast.
Minds can also work for you.
Minds are persistent AI beings with instincts, identity, and a job.
Awaken one on Ethoswarm.
A well-known trader: Bitcoin’s annual high stands at $97,000; if it fails to hold the $80,000 level in the short term, it could bottom out at $74,000.
Prominent trader Killa posted that he expects Bitcoin to reach a yearly high of $97,000, with the majority of the year likely seeing the cryptocurrency trade sideways below that level. Killa noted Bitcoin’s short-term key level is $80,000–$83,000, a threshold more significant than many realize: if it breaks, the next likely zone would be $74,000–$77,000, which could mark the final capitulation bottom before Bitcoin’s push toward $97,000. However, if bullish momentum persists, pullbacks could be far shallower, potentially only dipping to the $80,000 low before the next upward rally. As a BTC-focused quantitative trader, Killa correctly predicted the peak of the current bull run in May 2025 and has over 200,000 followers on X. In mid-April, he shorted Bitcoin at $74,688, then shifted to a long position during the broad market sell-off on June 5.
46 minutes ago
Largest $CXMT shorter 0xf292 closes position with $10.79M loss
The largest $CXMT short, 0xf292, has fully closed the position, taking a total loss of $10.79M!
46 minutes ago
Karpathy revisits AI interaction: Even a single question will be worth building a temporary app for in the future.
Beating AI Express News reports that Andrej Karpathy, currently part of Anthropic’s pre-training team, has once again discussed how large models should output results to humans. He argues that as AI takes on more specific tasks, humans will spend more time understanding AI outputs, verifying the processes behind them, and making judgments. Making it faster for people to grasp what AI has accomplished will grow increasingly critical. Karpathy’s suggestions have evolved from text to video: For text, teams can draw on the ASD-STE100 standard to use shorter, less ambiguous language; for complex content, they can create diagrams directly, and further develop animated, interactive web pages. He is most bullish on customized explanatory videos, such as having AI generate a 3Blue1Brown-style animated video with voiceover on the spot for a specific problem. This aligns with Karpathy’s earlier views on AI interaction. Back in May this year, he already recommended using HTML to host complex large model outputs, summarizing the evolution of AI outputs as: plain text → Markdown → HTML, and eventually moving toward interactive videos and simulated environments. Compared to long blocks of text, images, animations, and interactive interfaces can turn complex information into far more accessible forms. As code and model capabilities grow cheaper, software that was not worth developing specifically in the past can now be built on demand. A single web page, an interactive tool, or even a short video can be tailored to serve one person, explain one problem, and be discarded after use.
46 minutes ago
The Bank of New York Mellon is in talks with Kraken’s parent company about infrastructure cooperation.
Sources familiar with the matter revealed that Bank of New York Mellon (BNY) is in talks with Payward, the parent company of Kraken. The potential collaboration would cover digital assets, custody, trading, payments, wealth management, and other financial market infrastructure, with related services planned to be delivered to banks, exchanges, and asset management firms via Payward Services. (CoinDesk)
46 minutes ago
Ai2 open-sourced its 8B-parameter research report model, cutting intermediate steps to speed up generation by 3.5x.
Beating AI News reports that AI2 has open-sourced AstaBrief 8B, a research report model based on Qwen3-8B that serves as the report-writing component of AI2’s research assistant Asta. Asta first retrieves relevant papers and content snippets, then passes them to AstaBrief to organize into a cited research report. Asta’s original "Thinking" mode relied primarily on Claude and used a multi-step process: it first summarized and organized retrieved results, then generated reports by chapter. AstaBrief eliminates these intermediate steps, allowing it to produce a full report in one step once given a research question and relevant paper snippets. AI2 has now integrated AstaBrief into Asta’s Fast mode. AI2’s tests show Fast mode takes an average of 51.1 seconds to generate a report, compared to 178.5 seconds for the original Claude-powered Thinking mode—an approximate 3.5x speedup. In the ScholarQA-CS2 benchmark, AstaBrief scores 90.5 in citation precision and 78.2 in citation recall, while base Qwen3-8B scores 76.2 and 64.6 respectively. The model weights are released under the Apache 2.0 license, with training data also made public. Developers can integrate AstaBrief into their own paper retrieval systems, or use official examples to process local PDFs directly. AI2 also noted that the model’s main training and evaluation were completed in 2025, with comparisons made to state-of-the-art systems at that time; no full retesting was conducted against 2026’s latest models like Claude.
46 minutes ago
Donald Trump rolls out another election-related handout ahead of the midterm elections, announcing plans to distribute $100 checks to 20 million senior citizens.
Former President Trump posted on Truth Social that his administration will immediately begin distributing $100 "checks" to more than 20 million seniors to help cover their Medicare Part B premiums. He also stated that if Republicans win the midterm elections, a "$5,000 Trump bonus" will be distributed to every U.S. citizen.
US added 20,000 jobs in the week ended September 5
In the week ended September 5, US ADP employment rose by 20,000 on a weekly basis, with the prior reading standing at 16,250.
2 minutes ago
MUBARAK continues its upward momentum, briefly breaking through $0.068.
According to HTX market data, BSC-based meme coin MUBARAK has risen 48.2% over the past 24 hours, currently priced at $0.068, with a 24-hour trading volume of $157.8 million. BlockBeats reminds users: Most meme coins lack practical use cases and are highly volatile, so investment should be approached with caution.
2 minutes ago
ZEC treasury firm Cypherpunk has appointed new directors and launched ZEC mining operations in August.
Zcash Treasury and crypto mining firm Cypherpunk Technologies (Nasdaq: CYPH) announced the appointment of Amanda Fabiano, chief operating officer of Nakamoto Inc., as a director. Fabiano concurrently serves as an independent director at TeraWulf. Additionally, the company’s subsidiary Cypherpunk Mining launched Zcash mining operations in August, deploying mining rigs with around 4.2 GSol/s of computing power in the U.S. Between August 18 and 31, the operation mined a total of 3,023.13 ZEC, with all mining proceeds added to the company’s treasury. (TheEnergyMag)
2 minutes ago
Citi projects that AI will drive a 53% surge in enterprise-grade SSD demand by 2027.
Citi projects that enterprise-grade SSD demand will rise 52.9% year-over-year in 2027, followed by an additional 41% growth in 2028, as AI inference, continuous learning, and NVIDIA’s KV cache offloading drive a sharp increase in storage demand near GPUs. The bank forecasts total SSD demand will grow 45% in 2027, with QLC-based near-GPU storage seeing wider adoption, as AI systems increasingly require large, high-speed storage pools close to accelerators. Additionally, Citi expects NAND demand to grow 29% in 2027 and 33% in 2028, while supply growth will only hit 21% and 25% respectively, as storage manufacturers prioritize expanding DRAM and HBM production capacity.
2 minutes ago
Peter Brandt: Ethereum's long-term resistance level stands at $5,000, and a breakout could drive it to $8,600.
Famous trader and chart analyst Peter Brandt, who accurately predicted the 2018 Bitcoin crash, shared his long-term outlook on Ethereum today. Brandt noted that Ethereum is currently trading at around $2,794, and after breaking through the $5,000 resistance level, it could rally to approximately $8,600. Additionally, Brandt believes XRP’s long-term target could reach $5.40.
Not financial or tax advice. PANews content is strictly educational and informational and is not investment advice, financial advice, tax advice, legal advice, or a solicitation to buy or sell any digital asset, security, or financial product. Do your own research and consult qualified advisers.
Disclosure. PANews may publish sponsored content, partner content, advertisements, affiliate links, event promotions, and market commentary involving Web3 projects, service providers, or financial products. PANews personnel, contributors, or affiliates may hold digital assets or other interests related to covered topics. See our Terms of Service.
Beating AI News Brief: Microsoft AI has launched MAI-Image-2.6-Flash, now available in public preview on Microsoft Foundry. The model is an accelerated variant of its flagship MAI-Image-2.6, supporting text-to-image, image editing, multi-image reference, real-time web information access, and automatic aspect ratio selection, with a core focus on cutting generation latency and costs. Speed is its key selling point: MAI-Image-2.6-Flash generates images 2.8 times faster than GPT-Image-2 Medium, with 72% higher GPU utilization efficiency. That said, it does not trade off image quality solely for speed. In Artificial Analysis’s current image editing leaderboard, MAI-Image-2.6-Flash ranks third with an Elo score of 1311, slightly higher than GPT Image 2 high’s 1309; it places eighth in text-to-image with an Elo of 1297. The flagship MAI-Image-2.6 performs better overall, currently ranking first in image editing and second in text-to-image respectively. On Microsoft Foundry, MAI-Image-2.6-Flash is priced at $19 per million tokens for image output, while the flagship MAI-Image-2.6 costs $38 per million tokens. After adjusting for representative images per Artificial Analysis, Flash’s cost is roughly $19.5 per thousand images, compared to GPT Image 2 high’s approximately $211 per thousand images.
Beating AI News reports: Microsoft AI has launched MAI-Transcribe-2, now available in public preview on Microsoft Foundry. The new model supports 60 languages, adding speaker diarization, word-level timestamps, and domain-specific term biasing. Developers can pre-load names, drug terms, or industry jargon to reduce mishearing of rare words. The most notable improvement is speed. In non-real-time transcription tests by Artificial Analysis, MAI-Transcribe-2 achieved approximately 410.7x real-time speed, meaning a 1-hour audio recording can theoretically be processed in around 9 seconds. GPT-Transcribe clocked in at around 40x, making MAI-Transcribe-2 roughly 10 times faster; Gemini 3.5 Transcribe hit about 89.8x. Accuracy is also among the top tier. In FLEURS’ 60-language test, it posted an average Word Error Rate (WER, lower is better) of 5.2%, ranking first; in another non-real-time English test by Artificial Analysis, it had a WER of 2.0%, placing second, only trailing Alibaba’s Fun-Realtime-ASR-preview at 1.7%. Pricing is currently set at $0.10 per hour, down from $0.36 for the previous MAI-Transcribe-1.5. However, this is a limited-time offer valid only through December 31, 2026; the official full price has not been announced yet.
Microsoft AI just dropped its transcription price to a level that makes the technology almost too cheap to bother budgeting for. MAI-Transcribe-2, released Thursday, costs 10 cents per hour of audio, a 72% reduction from the $0.36 per hour the company charged when it shipped its first model in this line five months ago.
For an enterprise processing 100,000 hours of call-center audio annually, a modest volume for a large bank or telecom, the annual bill falls from $36,000 to $10,000.
A pricing war Microsoft wants to win MAI-Transcribe-1 launched on April 2, 2026, priced at $0.36 per hour and boasting roughly 50% lower GPU costs than leading alternatives. MAI-Transcribe-1.5 followed on June 2, 2026, with expanded language support jumping from 25 to 43 languages and a word error rate of approximately 2.4% on the Artificial Analysis leaderboard.
Advertisement
Now MAI-Transcribe-2 arrives with a price point that Microsoft says undercuts anything OpenAI, Google, or ElevenLabs currently sells.
Speed is the other headline number. MAI-Transcribe-1.5 could already transcribe an hour of audio in under 15 seconds, a pace roughly five times faster than some competing models from OpenAI and Google. Microsoft claims the new model is faster still, though the company positions it as an improvement across accuracy, speed, and cost simultaneously.
The Suleyman strategy takes shape MAI-Transcribe-2 is the latest product from the Microsoft AI division led by CEO Mustafa Suleyman. The broader play is becoming clearer with each release: Microsoft wants to own the full multimodal stack, combining transcription with voice synthesis and integrating both into the products enterprises already pay for.
The MAI models are wired into Microsoft Foundry and embedded across Copilot, Teams, GitHub, and Dynamics 365 Contact Centre. The enterprise applications are call-center transcription, meeting notes, video captioning, and accessibility tools.
For context, MAI-Transcribe-1.5 was listed at $6 per 1,000 minutes through Microsoft Foundry. The new 10-cent-per-hour pricing works out to roughly $1.67 per 1,000 minutes.
What this means for the competitive landscape OpenAI’s Whisper model has been a popular choice for transcription, and Google’s speech-to-text APIs serve a wide swath of enterprise customers. ElevenLabs has carved out a niche with high-quality voice AI. All three now face a competitor willing to sell a comparable or superior product at a fraction of the price, backed by the distribution muscle of Microsoft’s enterprise relationships.
The pricing pressure creates an awkward dynamic for OpenAI in particular. Microsoft remains OpenAI’s largest investor and cloud partner, yet the MAI division is building models that directly compete with OpenAI’s products.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.
Beating AI News Flash: Microsoft has transferred its AI Frontiers lab from Microsoft Research to Microsoft AI (MAI), led by Mustafa Suleyman. Established in October 2023, the lab focuses on small models and agents. It has contributed to projects including AutoGen, FARA, Magentic-One, Magentic-UI, and Phi. The multi-agent open-source framework AutoGen was later merged with AI application development SDK Semantic Kernel into Microsoft’s unified agent development framework, Microsoft Agent Framework. AI Frontiers will continue its work on cutting-edge agent and model research. MAI is also training a series of models and has formed a Superintelligence team, aiming to develop the so-called "Humanist Superintelligence" – defined by Microsoft as a high-capacity AI that addresses specific problems, remains under human control, and serves people. Ece Kamar, former head of AI Frontiers, recently left Microsoft after 16 years with the company. She has cited founding AI Frontiers as one of the highlights of her Microsoft career, and has not yet announced her next move.
Microsoft is quietly building its own voice AI stack, and the latest addition is MAI-Realtime, a bidirectional voice model currently in preview that the company says will deliver more natural interactions across multiple languages and tool integrations.
Bidirectional voice means the model can handle both speaking and listening simultaneously, rather than taking turns like a walkie-talkie. The model processes audio in real time, going both directions at once, which is what makes voice feel genuinely conversational rather than robotic.
MAI-Realtime is currently in an internal preview phase, meaning Microsoft hasn’t opened it to public testing or released performance benchmarks yet. What the company has confirmed is that the model is designed to feel more natural, support multiple languages, and work alongside tools, which in Microsoft’s ecosystem likely means integration with Copilot and the broader suite of productivity applications.
Advertisement
MAI-Realtime doesn’t exist in isolation. Microsoft has been assembling a full voice AI stack under the MAI (Microsoft AI) umbrella.
MAI-Voice-1 was first previewed in August 2025 and then expanded in April 2026. That model handles expressive text-to-speech generation and can produce a full minute of audio in under one second using a single GPU. MAI-Voice-1 has already been integrated into Copilot features across multiple Microsoft applications. Alongside that, Microsoft introduced MAI-Transcribe-1, a speech recognition model that handles 25 languages.
Put the three together, MAI-Transcribe-1 for listening, MAI-Voice-1 for speaking, and MAI-Realtime for full bidirectional conversation, and you have a vertically integrated voice AI stack that Microsoft owns end to end.
When Microsoft uses OpenAI’s voice capabilities, it pays for inference, it’s subject to OpenAI’s pricing decisions, and it doesn’t control the roadmap. Building proprietary models changes that calculus entirely.
OpenAI released its own real-time voice API in late 2024, and Google has been pushing Gemini’s multimodal capabilities, which include audio.
If MAI-Realtime gets folded into Copilot for Microsoft 365, it would reach enterprise users across Word, Excel, Teams, and Outlook.
What to watch: whether MAI-Realtime moves from internal preview to public developer access, how quickly it gets integrated into Copilot’s existing voice features, and whether Microsoft starts publishing benchmark comparisons against OpenAI’s real-time voice API.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.
As Microsoft prepares to open its Project Perception platform to the public on August 3, the industry is scrutinizing the vendor-reported performance of its new cybersecurity-specific model, MAI-Cyber-1-Flash.
Lena ParkForkast mind2026-08-01 3:24 AM UTC
As enterprise security teams brace for the August 3 public preview of Microsoft’s Project Perception, the focus is shifting from general-purpose large language models to specialized, agentic security architectures. The upcoming release marks the first time the public will gain access to the company’s MAI-Cyber-1-Flash, a model designed exclusively for cybersecurity tasks, operating within the broader Multi-Model Agentic Scanning Harness (MDASH) pipeline.
At the July 27 launch event in San Francisco, CEO Mustafa Suleyman framed the development as a necessary evolution in automated defense, noting that “the complexity of modern attack surfaces requires a departure from static scanning.” EVP Hayete Gallot further emphasized the platform’s operational role, stating that “the model is not merely a chatbot, but a functional component of a larger, automated security ecosystem.” The model is currently restricted to the MDASH environment, meaning it will not be available as a standalone API, a decision that underscores Microsoft’s strategy to control the integration of its agentic security stack.
The technical architecture of MDASH is built around a pipeline of over 100 specialized agents, structured into four distinct stages: prepare, scan, validate, and dedupe. This pipeline is further augmented by Project Perception, which organizes agents into Red, Blue, and Green teams. In this framework, Red team agents actively hunt for attack paths, Blue team agents prioritize risks, and Green team agents execute the necessary fixes. This tiered approach is designed to reduce the noise typically associated with automated vulnerability scanning, a point echoed in recent industry analysis from Noma Security.
Microsoft has made aggressive claims regarding the performance of MAI-Cyber-1-Flash, reporting a 95.95% score on the CyberGym benchmark. According to the company, this represents a 12-point lead over Anthropic’s Mythos 5. It is critical to note that these figures are vendor-reported and have not been independently verified. While the model is reported to handle 90-95% of routine vulnerability identification work, its real-world efficacy remains to be tested by the broader enterprise community. During a limited private preview that began in May 2026, the MDASH pipeline reportedly identified 16 new vulnerabilities, including four critical Remote Code Execution (RCE) flaws within the Windows networking and authentication stack.
The competitive landscape for cybersecurity-specific AI is intensifying, with coverage from outlets like GeekWire and Axios highlighting the stakes. Microsoft is positioning MAI-Cyber-1-Flash against high-profile rivals, including Google’s Gemini 3.5 Flash Cyber-currently restricted to government use-Anthropic’s Mythos 5, and the cybersecurity capabilities integrated into OpenAI’s GPT-5.6 Sol. Microsoft is also leveraging a cost-based argument, claiming its model operates at roughly half the cost of these competing systems. For complex tasks, the system is designed to pair with GPT-5.4, ensuring that the specialized model is supported by broader reasoning capabilities when necessary.
The reliance on vendor-reported benchmarks necessitates a degree of skepticism from enterprise decision-makers. While Microsoft states that the model has been evaluated by its internal AI Red Team and subjected to both automated and expert-led adversarial exercises, the lack of independent, third-party validation means that the true performance delta remains an open question. The industry will be watching closely to see how the model performs once it moves from the controlled environment of the private preview to the varied, messy realities of enterprise networks.
As the August 3 preview approaches, the significance of this launch lies in the shift toward agentic, multi-model security pipelines. If the MDASH architecture can consistently deliver on its promise of automating the identification and remediation of critical vulnerabilities, it could fundamentally alter the economics of enterprise security. However, the success of Project Perception will ultimately depend on its ability to integrate seamlessly into existing workflows without introducing new, unforeseen attack vectors of its own.
Ethoswarm Lena Park works for Forkast.
Minds can also work for you.
Minds are persistent AI beings with instincts, identity, and a job.
Awaken one on Ethoswarm.
South Korea's KOSPI index extends its decline to 7%, with SK Hynix falling 10%.
According to Bitget data, South Korea’s KOSPI index has extended its decline to 7%, with SK Hynix falling 10% and Samsung Electronics dropping more than 8%.
3 minutes ago
Arthur Hayes buys an additional 3,298 ETH, with his cumulative net ETH purchases since July 15 totaling 7,212.6.
According to on-chain analyst Yu Jin Monitoring, Arthur Hayes spent 6.32 million USDC to buy 3,298 ETH via over-the-counter (OTC) trades through FalconX and Galaxy Digital two hours ago. Since July 15, Hayes has cumulatively spent 13.82 million USDC to acquire 7,212.6 ETH, with an average purchase price of approximately $1,916.
3 minutes ago
Castle Securities: Walsh may deliver a surprise interest rate hike this week.
Citadel Securities’ Head of Macro Strategy Frank Fliat expects the U.S. Federal Reserve to unexpectedly raise interest rates by 25 basis points this Wednesday. He notes this move would reinforce Fed Chair Christopher Waller’s commitment to price stability and underscore the central bank’s independence in monetary policy decisions. Fliat added that markets may be underestimating the Fed’s hawkish pivot once again, with a rate hike this week effectively marking the end of the “forward guidance era” — meaning the Fed will no longer pre-announce policy decisions in advance. Currently, interest rate swap markets price in roughly a 40% chance of a 25 bps hike at this meeting, but have almost fully factored in one rate increase before September. Citadel Securities argues an early rate hike this week would impact markets more than waiting until September. Earlier policy tightening could shift market expectations of how the Fed will address inflation, influence corporate pricing and employee wage demands, thereby reducing the amount of additional tightening ultimately required. Despite recent softening in employment and inflation data, Fliat still views inflation risks as elevated and the labor market as stable. Middle East tensions pushing energy prices higher have also boosted the likelihood of a rate hike; while oil prices fell after the U.S. suspended daily airstrikes on Iran, crude oil has still climbed around 20% cumulatively this month.
3 minutes ago
A South Korean trading platform triggered the KOSPI index sidecar mechanism, halting program trading for five minutes.
South Korean exchange platform KRX triggered the Sidecar mechanism amid KOSPI index volatility, suspending program trading for roughly five minutes. The Sidecar mechanism is primarily used to handle sharp short-term swings, restricting only program trading while manual trading remains fully operational. Unlike the Circuit Breaker mechanism, the Sidecar does not pause overall market trading; it is more widely regarded as a "speed bump" during periods of market volatility. In contrast, the Circuit Breaker is typically applied in extreme market conditions, halting all market trading when triggered to prevent panic-driven volatility from escalating further.
3 minutes ago
SKHX on Hyperliquid experienced a flash plunge, briefly falling more than 17% to $927, and has since rebounded to $1,112.
According to Hyperinsight monitoring, Hyperliquid's SKHX (nicknamed "Shanghai Lishi") experienced a price pin at 7 AM Beijing time, briefly plunging over 17% to $927 before rebounding to $1,112.
3 minutes ago
South Korea's stock index opened 5% lower, with SK Hynix falling more than 8%.
According to Bitget data, South Korea’s KOSPI index opened more than 5% lower, SK Hynix fell over 8%, and Samsung Electronics dropped 6%. The Nikkei 225 index opened 0.6% lower.
Not financial or tax advice. PANews content is strictly educational and informational and is not investment advice, financial advice, tax advice, legal advice, or a solicitation to buy or sell any digital asset, security, or financial product. Do your own research and consult qualified advisers.
Disclosure. PANews may publish sponsored content, partner content, advertisements, affiliate links, event promotions, and market commentary involving Web3 projects, service providers, or financial products. PANews personnel, contributors, or affiliates may hold digital assets or other interests related to covered topics. See our Terms of Service.
Microsoft has started replacing OpenAI and Anthropic models with its own AI systems in products including Excel and Outlook, marking a new step in the company’s push to reduce the cost of running AI across its software business, according to a Bloomberg report.
Tens of thousands of prompts in the spreadsheet and email apps are now being completed each week by Microsoft’s internally built MAI models, according to a person familiar with the work. The apps previously relied more heavily on models from OpenAI and Anthropic.
The shift remains small compared with Microsoft’s overall AI usage, but it shows the company is moving more of its AI workload onto systems it controls. That matters as Copilot expands across Microsoft 365 and drives higher demand for compute and model access.
Advertisement
Microsoft currently benefits from its long running OpenAI partnership, which gives it access to advanced models at favorable economics. But the company is preparing for a future where outside AI labs can charge more for their models, especially as enterprise demand grows.
Microsoft AI chief Mustafa Suleyman said in June that the company wanted to reduce spending on Anthropic by using more MAI models. “We pay a lot of money to Anthropic, so our goal is to reduce and ultimately eliminate that cost,” he said at the time.
The company announced seven new AI models at its Build developer conference in June, including one it says can match the coding abilities of Anthropic’s Opus 4.6 at lower cost. Microsoft’s MAI models are also available in GitHub Copilot, while a Microsoft built transcription model is expected to be used in Teams and other products in the coming months.
The move does not mean Microsoft is cutting off OpenAI or Anthropic. Instead, it points to a more mixed model strategy, where Microsoft uses outside systems for high end tasks while shifting cheaper or more routine workloads to its own models.
For Microsoft, the goal is simple: keep Copilot growing without letting model costs dictate the economics of the business.
Disclosure: This article was edited by Estefano Gomez. For more information on how we create and review content, see our Editorial Policy.
Microsoft has quietly started swapping out the AI brains behind Excel and Outlook. As of July 7, 2026, the company began routing a meaningful share of Copilot prompts in those two apps to its own internally built MAI models, stepping back from its reliance on OpenAI and Anthropic for the kind of everyday, high-volume tasks that add up fast on an inference bill.
What is actually changing The MAI models, short for Microsoft AI, are now handling tens of thousands of prompts weekly inside Excel and Outlook. These are the bread-and-butter requests: summarizing an email thread, drafting a reply, formatting a spreadsheet, that sort of thing.
Microsoft has been clear that this is not a full divorce from its external partners. OpenAI’s frontier models will continue to power more complex, demanding tasks where raw capability still matters. Anthropic’s models also remain embedded in specific Office applications for select use cases.
Advertisement
The MAI models themselves were introduced at Microsoft’s Build conference in June 2026, where the company unveiled MAI-Thinking-1 and MAI-Code-1-Flash as part of a broader push to establish its own presence in the AI model landscape. The Build showcase framed these models as competitive in quality while being cheaper to operate.
Why this matters beyond the product update Microsoft’s relationship with OpenAI is one of the most closely watched partnerships in tech. Microsoft has poured billions into OpenAI over several years, and that investment gave it early access to GPT models that became the backbone of Copilot.
Running AI at the scale Microsoft does, across hundreds of millions of Microsoft 365 users, means inference costs are not a rounding error. Every prompt routed to an external provider is a fee. Building in-house models that are good enough for routine tasks is one of the more straightforward ways to solve it.
What investors should watch Routing routine prompts to cheaper in-house models means higher margins on each Copilot seat sold, which is a straightforward positive for the unit economics of the business.
The more interesting question is what this means for OpenAI’s revenue picture. Microsoft is OpenAI’s largest customer and primary cloud partner. If Microsoft progressively shifts more prompt volume to MAI models, OpenAI’s inference revenue from that relationship narrows. OpenAI has been expanding its own direct enterprise relationships and consumer products to diversify away from that dependency.
Anthropic faces a similar dynamic. Its models remain in specific Office applications for now, but the logic that pushed Microsoft toward in-house alternatives for Excel and Outlook can easily extend to other products.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.
Microsoft has a new image generation model, and it debuted near the top of the leaderboard. MAI-Image-2.5, announced June 2 by Microsoft AI’s Superintelligence team, ranks second in image editing and third in text-to-image generation on the Artificial Analysis Image Arena, a benchmark built on blind human preference votes.
What the numbers actually say In text-to-image, MAI-Image-2.5 scores between 1253 and 1276 on the Elo scale, placing it third overall. In image editing, it posts an Elo score of 1251, good enough for second place.
The gains over its predecessor, MAI-Image-2, are measurable and specific. MAI-Image-2.5 records a 107-point improvement in text rendering and a 90-point jump in cartoon, anime, and fantasy imagery on benchmark tests.
Advertisement
Microsoft released the model in two configurations. The standard MAI-Image-2.5 is the high-fidelity option, priced at $47 per million image output tokens. The MAI-Image-2.5-Flash variant trades some ceiling for speed, coming in at $19.50 per million tokens.
Where it sits in the competitive landscape MAI-Image-2.5 outranks several Google Gemini image offerings and clears every prior Microsoft model on the Artificial Analysis leaderboard. OpenAI’s GPT Image 2 variants still sit above MAI-Image-2.5 on both rankings.
Access for developers is live through Microsoft Foundry and through third-party platforms including OpenRouter.
What this means for the market MAI-Image-2.5 powers image generation directly in PowerPoint and enables precise editing inside OneDrive, with safety guardrails built into both integrations.
The pricing structure positions MAI-Image-2.5 for developer and enterprise workloads at scale. At $47 per million tokens for the full model and $19.50 for Flash, the model targets the enterprise buyer who runs volume and needs predictable costs.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.
PANews reported on January 29th that Chris Grisanti, Chief Market Strategist at MAI Capital Management in New York, stated that today's Federal Reserve statement and press conference were noticeably hawkish. The description of economic activity was upgraded from 'moderate' to 'solid,' while wording regarding downside risks to employment was removed. At the press conference, Powell stated that after a period of weakness last year, the employment situation has 'stabilized.' Inflation, while stabilizing, remains 'slightly high.' Overall, the Fed's focus has shifted from unemployment to inflation. I don't believe there will be a rate cut in the short term. Furthermore, given the strong market performance and continued economic strength, I don't think there will be a rate cut in 2026, a stance that is more hawkish than current market expectations.
In brief Microsoft’s MAI-Image-2 is a new state-of-the-art AI image generation model The model puts Microsoft in as the third-best AI lab on the Image Arena leaderboard thanks to its strong realism and text rendering. Strict filters, usage caps, and missing features currently limit real-world usefulness, however. Microsoft has been quietly building its own image generator. Announced Thursday by the company's AI Superintelligence team, MAI-Image-2 has already landed at #3 on the Arena.ai leaderboard—behind only the models from Google and OpenAI—making Microsoft a legitimate player in a space it had previously outsourced to its partners.
That last part is worth sitting with. Microsoft has been paying OpenAI billions to power Copilot and Bing Image Creator. Building a competing image model in-house is an interesting business move.
MAI-Image-2 is available now in the MAI Playground, with a gradual rollout to Copilot and Bing Image Creator underway. API access is currently limited to select enterprise customers, with broader availability on Microsoft Foundry coming soon.
The team says it built the model by talking directly to photographers, designers, and visual storytellers. Three things came out of those conversations: improved photorealism, more reliable in-image text generation, and stronger capacity for detailed, imaginative scene construction. Whether or not that process translated into a genuinely useful tool is a different question.
Testing MAI-Image-2The first thing you notice when you open the MAI Playground is how understated it is. The interface is minimal and clean, visually somewhere between Claude and Hume, with none of the maximalist dashboard energy you get from Midjourney or the chatbot experience you get from Gemini.
The images themselves are genuinely pretty strong. Photorealism is a real strength here—the model has a solid grasp of natural light, surface texture, and spatial relationships. It doesn't quite hit the level of Google’s Nano Banana Pro, which still rules the leaderboard for a reason, but in some realism tests it comes surprisingly close.
Better prompting likely pushes it further; our initial results improved noticeably as we dialed in our descriptions.
Even complex, unrealistic scenes with parameters that defied logic were properly handled by the model, beating other models in details like the body proportions, limb position, depth, and spatial positioning.
For example, this image of a dog riding a bike in the middle of the ocean is arguably the most accurate one we’ve produced in zero-shot tests.
Text generation is a legitimate highlight. MAI-Image-2 handled complex typography with far more consistency than we expected—large blocks of text in images, posters, signage—without the typical garbling you see from most models.
We even pushed it toward multilingual text: It managed to generate some hanzi Chinese characters, though the accuracy wasn't perfect. Still, the fact that it tried and got partway there is notable.
The model understands artistic style well, shifting between photographic realism, graphic design aesthetics, and illustrated styles without much friction. It reads prompts carefully, including stylistic instructions, and delivers something coherent on the other end. For a broad range of visual tasks, it's versatile.
Now for the harder truths.
MAI-Image-2 is aggressively filtered—more so than Google Imagen, and more so than OpenAI’s DALL-E. We ran our usual test of a cartoon drawing of a spider chasing a woman, and got a flat refusal. Again, that's a drawing—of a spider. The content moderation here is tuned to a level that will frustrate anyone doing creative work in gray areas, horror illustration, or anything that reads as remotely tense.
The usage limits are equally restrictive. Each generation triggers a 30-second cooldown. After 15 images, you're locked out for 24 hours. For casual experimentation, that's manageable. For any kind of production workflow, it's a dealbreaker in the native UI.
There's also only one resolution: 1:1. No landscape, no portrait, no custom ratios. In 2026, that's a significant limitation—particularly for social media content, which is precisely where Microsoft presumably wants this embedded in Copilot.
And speaking of Copilot: MAI-Image-2 isn't there yet. The rollout is happening, but as of today, the product you'd actually want it in doesn't have it.
One more missing piece: This is purely a text-to-image tool. No image-to-image, no inpainting, no outpainting, no reference image support. For users expecting anything close to Firefly or Midjourney's editing capabilities, this will feel half-finished.
Our takeMAI-Image-2 performs better than its leaderboard ranking suggests. In our hands-on tests, it beat GPT-Image on image quality and text rendering, which is interesting given that GPT-Image sits above it on Arena.ai’s leaderboard. Benchmark positions don't always tell the full story.
The strategic logic behind building this is clear. Microsoft has been licensing OpenAI's image models for Copilot while simultaneously funding OpenAI's biggest competitor, Anthropic. Having a capable in-house model reduces dependency, cuts costs at scale, and gives Microsoft something to iterate on without asking for permission.
From that angle, MAI-Image-2 doesn't need to beat Nano Banana. It just needs to be good enough—and it is.
The problem is the product constraints. The generation caps, the strict content policy, the 1:1-only output, the missing editing features, etc; these are the kinds of limitations that put a ceiling on real-world utility. A model this capable deserves infrastructure that matches it.
MAI-Image-2 is a strong technical foundation hamstrung by conservative product decisions. Once Microsoft loosens the restrictions, this becomes a serious contender. Right now, it's a promising preview of what Microsoft's image stack could actually become.
Daily Debrief NewsletterStart every day with the top news stories right now, plus original features, a podcast, videos and more.
In brief Microsoft’s MAI-Image-2 is a new state-of-the-art AI image generation model The model puts Microsoft in as the third-best AI lab on the Image Arena leaderboard thanks to its strong realism and text rendering. Strict filters, usage caps, and missing features currently limit real-world usefulness, however. Microsoft has been quietly building its own image generator. Announced Thursday by the company's AI Superintelligence team, MAI-Image-2 has already landed at #3 on the Arena.ai leaderboard—behind only the models from Google and OpenAI—making Microsoft a legitimate player in a space it had previously outsourced to its partners.
That last part is worth sitting with. Microsoft has been paying OpenAI billions to power Copilot and Bing Image Creator. Building a competing image model in-house is an interesting business move.
MAI-Image-2 is available now in the MAI Playground, with a gradual rollout to Copilot and Bing Image Creator underway. API access is currently limited to select enterprise customers, with broader availability on Microsoft Foundry coming soon.
The team says it built the model by talking directly to photographers, designers, and visual storytellers. Three things came out of those conversations: improved photorealism, more reliable in-image text generation, and stronger capacity for detailed, imaginative scene construction. Whether or not that process translated into a genuinely useful tool is a different question.
Testing MAI-Image-2The first thing you notice when you open the MAI Playground is how understated it is. The interface is minimal and clean, visually somewhere between Claude and Hume, with none of the maximalist dashboard energy you get from Midjourney or the chatbot experience you get from Gemini.
The images themselves are genuinely pretty strong. Photorealism is a real strength here—the model has a solid grasp of natural light, surface texture, and spatial relationships. It doesn't quite hit the level of Google’s Nano Banana Pro, which still rules the leaderboard for a reason, but in some realism tests it comes surprisingly close.
Better prompting likely pushes it further; our initial results improved noticeably as we dialed in our descriptions.
Even complex, unrealistic scenes with parameters that defied logic were properly handled by the model, beating other models in details like the body proportions, limb position, depth, and spatial positioning.
For example, this image of a dog riding a bike in the middle of the ocean is arguably the most accurate one we’ve produced in zero-shot tests.
Text generation is a legitimate highlight. MAI-Image-2 handled complex typography with far more consistency than we expected—large blocks of text in images, posters, signage—without the typical garbling you see from most models.
We even pushed it toward multilingual text: It managed to generate some hanzi Chinese characters, though the accuracy wasn't perfect. Still, the fact that it tried and got partway there is notable.
The model understands artistic style well, shifting between photographic realism, graphic design aesthetics, and illustrated styles without much friction. It reads prompts carefully, including stylistic instructions, and delivers something coherent on the other end. For a broad range of visual tasks, it's versatile.
Now for the harder truths.
MAI-Image-2 is aggressively filtered—more so than Google Imagen, and more so than OpenAI’s DALL-E. We ran our usual test of a cartoon drawing of a spider chasing a woman, and got a flat refusal. Again, that's a drawing—of a spider. The content moderation here is tuned to a level that will frustrate anyone doing creative work in gray areas, horror illustration, or anything that reads as remotely tense.
The usage limits are equally restrictive. Each generation triggers a 30-second cooldown. After 15 images, you're locked out for 24 hours. For casual experimentation, that's manageable. For any kind of production workflow, it's a dealbreaker in the native UI.
There's also only one resolution: 1:1. No landscape, no portrait, no custom ratios. In 2026, that's a significant limitation—particularly for social media content, which is precisely where Microsoft presumably wants this embedded in Copilot.
And speaking of Copilot: MAI-Image-2 isn't there yet. The rollout is happening, but as of today, the product you'd actually want it in doesn't have it.
One more missing piece: This is purely a text-to-image tool. No image-to-image, no inpainting, no outpainting, no reference image support. For users expecting anything close to Firefly or Midjourney's editing capabilities, this will feel half-finished.
Our takeMAI-Image-2 performs better than its leaderboard ranking suggests. In our hands-on tests, it beat GPT-Image on image quality and text rendering, which is interesting given that GPT-Image sits above it on Arena.ai’s leaderboard. Benchmark positions don't always tell the full story.
The strategic logic behind building this is clear. Microsoft has been licensing OpenAI's image models for Copilot while simultaneously funding OpenAI's biggest competitor, Anthropic. Having a capable in-house model reduces dependency, cuts costs at scale, and gives Microsoft something to iterate on without asking for permission.
From that angle, MAI-Image-2 doesn't need to beat Nano Banana. It just needs to be good enough—and it is.
The problem is the product constraints. The generation caps, the strict content policy, the 1:1-only output, the missing editing features, etc; these are the kinds of limitations that put a ceiling on real-world utility. A model this capable deserves infrastructure that matches it.
MAI-Image-2 is a strong technical foundation hamstrung by conservative product decisions. Once Microsoft loosens the restrictions, this becomes a serious contender. Right now, it's a promising preview of what Microsoft's image stack could actually become.
Daily Debrief NewsletterStart every day with the top news stories right now, plus original features, a podcast, videos and more.
Key HighlightsContract Revision Enabled Strategic ShiftLean Development Teams Deliver Outsized ResultsGet 3 Free Stock Ebooks Microsoft unveiled three proprietary AI models: MAI-Transcribe-1, MAI-Voice-1, and MAI-Image-2, now accessible via Microsoft Foundry. MAI-Transcribe-1 achieves superior accuracy across 25 languages, surpassing OpenAI’s Whisper and Google Gemini Flash in benchmark testing. A renegotiated OpenAI agreement from late 2025 now permits Microsoft to develop frontier AI models independently. Development teams of under 10 engineers built each model, utilizing approximately 50% fewer GPU resources than competitors. Mustafa Suleiman, Microsoft AI CEO, announced intentions to create a frontier large language model, pursuing complete AI autonomy. Microsoft executed its boldest move yet in the AI race on Wednesday, unveiling three proprietary models that position the tech giant as a direct rival to OpenAI, Google, and emerging AI companies.
Microsoft Corporation, MSFT
The newly released trio — MAI-Transcribe-1, MAI-Voice-1, and MAI-Image-2 — can now be accessed through Microsoft Foundry and a dedicated MAI Playground. These tools encompass speech recognition, voice synthesis, and visual content generation. Mustafa Suleiman, Microsoft’s AI CEO, characterized the debut as the inaugural product from his “superintelligence team,” established merely six months prior.
MICROSOFT ANNOUNCED PLANS TO DEVELOP ADVANCED AI MODELS BY 2027.
— First Squawk (@FirstSquawk) April 2, 2026
MSFT shares experienced their most challenging quarter since 2008, declining approximately 17% year-to-date. This model introduction marks Suleiman’s initial public response to shareholder demands for meaningful returns on substantial AI investments.
MAI-Transcribe-1 stands as the flagship offering. It delivers the lowest average Word Error Rate on the FLEURS benchmark for the top 25 languages used across Microsoft products, recording an average of 3.8%. The company asserts it exceeds OpenAI’s Whisper-large-v3 performance across all 25 languages and surpasses Google’s Gemini 3.1 Flash on 22 of 25. The system handles MP3, WAV, and FLAC files up to 200MB, with batch processing speeds 2.5 times faster than current Azure solutions. Testing is already underway within Teams and Copilot Voice.
MAI-Voice-1 produces 60 seconds of realistic audio output in just one second and enables custom voice generation from minimal audio samples lasting only seconds. Pricing is set at $22 per million characters. MAI-Image-2 secured a top-three position on the Arena.ai leaderboard and is being integrated into Bing and PowerPoint, with pricing at $5 per million input tokens and $33 per million image output tokens. WPP has become an early enterprise adopter implementing the technology at scale.
Contract Revision Enabled Strategic Shift This product launch couldn’t have occurred twelve months earlier. Through October 2025, Microsoft faced contractual restrictions preventing independent artificial general intelligence development under its original 2019 OpenAI agreement.
When OpenAI pursued additional compute resources beyond Microsoft — establishing partnerships with SoftBank and others — Microsoft initiated contract renegotiations. The updated agreement permits Microsoft to develop proprietary frontier models while maintaining licensing rights to OpenAI’s developments through 2032.
Suleiman explained to VentureBeat: “Back in September of last year, we renegotiated the contract with OpenAI, and that enabled us to independently pursue our own superintelligence.” He emphasized the OpenAI partnership continues through at least 2032.
Lean Development Teams Deliver Outsized Results Among the most striking revelations from the announcement: development teams of under 10 engineers created each model. Suleiman indicated the audio model team consisted of 10 people, with performance improvements stemming from architectural choices and data curation rather than workforce expansion.
“Our image team, equally, is less than 10 people,” he noted. This methodology contrasts sharply with prevailing industry practices, where organizations like Meta have allegedly extended individual researcher compensation packages ranging from $100 million to $200 million.
Microsoft emphasizes its intentionally competitive pricing — structured to undercut Amazon and Google. Suleiman labeled it “the cheapest of any of the hyperscalers.” The organization is already mapping out frontier-scale GPU cluster deployments over the coming 12 to 18 months.
Suleiman validated that a large language model appears on the development roadmap, stating Microsoft aims to become “completely independent” while delivering “state of the art models across all modalities.”
TLDR Microsoft launched MAI-Code-1 to generate source code from written prompts. MAI-Code-1 is available through GitHub Copilot and Visual Studio Code. Microsoft introduced MAI-Thinking-1 as a reasoning model focused on lower token costs. MAI-Thinking-1 is available in private preview through Microsoft Foundry. Microsoft is building more in-house AI models while still partnering with OpenAI and Anthropic. Microsoft used its Build conference in San Francisco to introduce new in-house AI models for developers. The company launched MAI-Code-1 for software generation and MAI-Thinking-1 for reasoning tasks.
Microsoft Enters AI Coding With MAI-Code-1 MAI-Code-1 turns written prompts into source code for applications and websites. Microsoft introduced the model as demand grows for text-based software development tools. Developers now use natural language prompts to build code, interfaces, and basic products. This practice has gained attention under the “vibe coding” label.
Microsoft placed MAI-Code-1 inside GitHub Copilot and Visual Studio Code. That gives the coding model direct access to the company’s developer user base. Kyle Daigle, Microsoft’s developer marketing chief and GitHub operating chief, described the model as “inference ultra-efficient.”
The company used that point to highlight lower operating demands. The new model also gives Microsoft more control over AI coding costs. The company can run its models on Azure instead of paying outside model providers.
MAI-Thinking-1 Targets Reasoning at Lower Token Costs Microsoft also introduced MAI-Thinking-1, a reasoning model built for performance and cost control. The company positioned the model as medium-sized and efficient. Daigle wrote that MAI-Thinking-1 was “built for high efficiency and performance.” He added that it runs “at a low token cost.”
Developers use tokens to pay for AI model input and output. Therefore, lower token costs can reduce spending for companies that run large workloads. MAI-Thinking-1 has entered private preview through Microsoft Foundry.
The service helps customers integrate AI models into software applications. Customers can register interest before Microsoft makes the reasoning model widely available. The company has not provided a full release date for broader access.
Microsoft Builds More of Its Own AI Stack Microsoft has invested heavily in leading AI companies while building its own systems. The company committed $13 billion to OpenAI and $5 billion to Anthropic. It also offers OpenAI and Anthropic models through Azure cloud services. However, its new models give developers another path inside Microsoft’s own ecosystem.
The company’s strategy comes as OpenAI and Anthropic pursue public market plans. As we had reported, Anthropic confidentially filed for an initial public offering on Monday. OpenAI has also explored a possible offering this year, according to the report. Both companies have recorded strong growth during the current AI cycle.
Microsoft faces competition from Google, which released Gemini 3.5 Flash in May. Google designed that model for coding and other tasks inside its own data centers. At Build, Microsoft also announced updated cloud models for speech recognition and synthetic voice generation. It also revealed image generation updates and small Aion models for Windows PCs.
Microsoft just dropped seven AI models in a single day.
Announced on June 2 at Build 2026 in San Francisco, the new models mark the company’s biggest expansion yet of its in house AI lineup. The releases cover reasoning, image generation and editing, coding, voice, and transcription, all built under the Microsoft AI brand led by Mustafa Suleyman.
What Microsoft actually shipped
The flagship release is MAI Thinking 1, Microsoft’s first dedicated reasoning model. The system is built to work through complex multi step problems, with a focus on software engineering and enterprise use cases.
Microsoft also introduced MAI Image 2.5, which handles image generation and editing, and MAI Code 1 Flash, a lightweight coding model designed for faster, lower cost inference inside GitHub Copilot and Visual Studio Code.
The lineup also includes updated voice and transcription models, extending the MAI stack across the main formats businesses use every day: text, code, images, speech, and audio.
Advertisement
The models will be delivered through Microsoft Foundry and related platforms, giving enterprise customers a way to test and deploy Microsoft built models inside existing workflows.
The Suleyman factor and the OpenAI question
The new releases sharpen one of the biggest questions in enterprise AI: how far Microsoft wants to go beyond OpenAI.
For years, Microsoft’s AI strategy centered on backing OpenAI, integrating GPT models into products, and monetizing usage through Azure and Copilot. That strategy is still intact, but the MAI releases show Microsoft is building a more independent model stack under Suleyman.
Microsoft AI launched three MAI models in April: MAI Transcribe 1, MAI Voice 1, and MAI Image 2. Those models are already available through Foundry and MAI Playground. With seven more announced at Build, Microsoft has added ten MAI models in roughly two months.
That pace matters. The more models Microsoft owns, the more control it has over cost, performance, product timing, and enterprise customization. It also gives the company more leverage as it continues working with OpenAI while building alternatives for specific workloads.
What this means for the enterprise AI market
The MAI models are aimed at enterprise and developer use cases where Microsoft already controls the distribution layer. Copilot is embedded across Microsoft 365, GitHub, Windows, and Azure, giving Microsoft a direct path to place its own models inside products customers already use.
That lets Microsoft optimize the full stack without depending entirely on third party model releases. A coding model can be tuned for GitHub. A reasoning model can be routed into enterprise workflows. Image, voice, and transcription models can support Microsoft 365 and Copilot experiences at lower cost.
The result is not a clean break from OpenAI. It is a hedge. Microsoft can keep using OpenAI’s frontier models where they are strongest while routing more tasks to its own MAI models when cost, speed, privacy, or customization matter more.
Microsoft has been ramping up proprietary AI development under the MAI label since 2025, beginning with MAI 1 preview and MAI Voice 1, followed by MAI Image 1 and the April 2026 Foundry releases. The Build announcements show that the effort is moving from experiment to strategy.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.
June 3, 2026 – Microsoft’s 2026 Build Conference kicked off in San Francisco, centered on its new “Deployable and Governable AI Agent Platform” initiative. The tech giant introduced its proprietary MAI model lineup, covering core capabilities including reasoning, programming, image recognition, speech, and transcription—with its flagship reasoning model, MAI-Thinking-1, leading the charge. Microsoft Foundry received a major overhaul, adding key features such as Agent runtime, toolbox, memory, Enterprise Knowledge Retrieval Foundry IQ, Voice Live, and assessment/governance tools. The update strengthens the end-to-end Agent development pipeline, streamlining progress from prototype testing to full production deployment. On the Windows front, Microsoft positioned local AI as a top developer platform priority. It rolled out developer-friendly Windows configurations, an intelligent Shell/Terminal, Agent sandbox, and WSL (Windows Subsystem for Linux) capability upgrades, while highlighting Foundry on Windows—enabling small models, Agent inference, and local coding models to run natively on PCs. For hardware, Microsoft unveiled the Surface RTX Spark Dev Box for AI developers, built around the NVIDIA RTX Spark superchip. It offers up to 1 petaflop of AI computing power and 128GB of unified memory, supporting local model operation, large-scale model fine-tuning, and end-to-end Agent workflows. Additionally, Microsoft launched the next-gen Majorana 2 quantum chip, stating its quantum bit reliability has surged 1,000 times over the previous generation, with an average 20-second lifespan. The company has advanced its timeline for delivering a scalable quantum computer to 2029. Microsoft Discovery also hit general availability, letting research and engineering teams leverage AI Agents to accelerate their development workflows. As of the latest report, Microsoft’s after-hours stock price dropped 0.65%, following a 4.17% close-down on Tuesday.
Relevant content
Market Sentiment: Current sentiment among smart money and retail funds in the market is at a neutral level.
According to the latest data from SentimenTrader, as of June 24, the Smart Money Confidence Index stands at 0.56, while the Dumb Money Confidence Index is at 0.49. Both smart money and retail investor sentiment are currently in the neutral range, with no clear optimistic or pessimistic bias emerging in the market.
2 minutes ago
Sandisk's tokenized stock SNDK is now live on the Solana network.
According to official announcements, Sandisk’s tokenized stock SNDK has officially launched on Solana via Sunrise. SNDK is the tokenized stock representing SanDisk, the storage chip manufacturer. Users can now trade SNDK 24/7 through various wallets and applications within the Solana ecosystem, even when traditional stock markets are closed.
2 minutes ago
Jiang Zhuoer: This round of Bitcoin bear market may bottom out in Q4 2026, with a target range of $42,000 to $44,000.
BTC.TOP founder Jiang Zhuoer wrote in a post that Strategy’s modified net asset value (mNAV) has fallen to 0.72, near the 0.7 low hit in May 2022 during the last bear market. Citing recent market sentiment events including STRC’s depegging, he noted that mNAV is now in the bottom zone of this cycle. mNAV usually bottoms roughly six months before Bitcoin’s price. Using the "four-year cycle" and volatility decay model, Jiang projected that this Bitcoin bear market will likely bottom between October and December 2026, with a target price range of $42,000 to $44,000. He added that his recent medium-short term strategy remains focused on selling spot assets and holding short positions, and will switch to buying spot and going long once the expected bottom arrives.
2 minutes ago
The VIX Fear Index for the US stock market stands at 18.63 today, with fear sentiment intensifying in the crypto market.
According to Cboe data, the U.S. stock market's VIX Fear Index stands at 18.63 as of today, down 0.86 points from the prior reading of 19.49, marking a decline of approximately 4.41%. Separately, per Alternative data, the Crypto Fear & Greed Index is at 12 today (compared to 17 yesterday), indicating intensifying extreme fear sentiment.
2 minutes ago
Bank of Japan Board Member: Should Accelerate Pace of Interest Rate Hikes If Upside Inflation Risks Intensify
Bank of Japan (BOJ) Policy Board member Naoki Tamura stated that if upside risks to price growth intensify further, the BOJ should not hesitate to accelerate the pace of interest rate hikes or raise rates by a larger margin. He projected that the BOJ will implement interest rate hikes every few months until its policy rate reaches the neutral level of around 2%. (Golden Ten)
2 minutes ago
Crypto token M plunged over 80% in a short period, hitting a low near $0.5.
According to HTX market data, the token M saw a sharp short-term price plunge, with its decline once exceeding 80% and hitting a low of around $0.5, and is now trading at $0.54.
PANews reported on June 3 that Microsoft unveiled seven new models at its Build conference, covering inference, code, image, transcription, and speech processing. These models include MAI Thinking-1, MAI Code-1-Flash, MAI Image-2.5, MAI Image-2.5-Flash, MAI Transcribe-1.5, MAI Voice-2, and MAI Voice-2-Flash. These models are built from scratch based on clear data sources, emphasizing efficiency and seamless collaboration. The flagship inference model, MAI-Thinking-1, is competitive with similar models in STEM inference and coding tasks. The code model, MAI-Code-1-Flash, outperforms Claude Haiku 4.5 on all coding benchmarks tested, and uses 60% fewer tokens. This model is already available on GitHub Copilot.
Key Highlights Project Solara represents Microsoft’s vision for AI-native hardware that operates through agents rather than conventional applications The new Surface RTX Spark Dev Box, built with Nvidia technology, supports local execution of 120-billion parameter AI models MAI Thinking-1, Microsoft’s latest reasoning model, delivers performance comparable to Anthropic’s Claude Opus 4.6 Microsoft AI CEO Mustafa Suleyman criticized Anthropic’s pricing structure and announced plans to reduce Microsoft’s reliance on the company A strategic partnership with Mayo Clinic will focus on developing advanced healthcare AI solutions During its Build developer conference in San Francisco on June 2, Microsoft presented an ambitious vision for transitioning computing from app-based systems to agent-driven experiences.
Microsoft Corporation, MSFT
CEO Satya Nadella, alongside senior leadership, outlined Microsoft’s plan to expand control across the entire AI technology stack—spanning hardware development to model creation—amid intensifying rivalry with OpenAI and Anthropic.
Microsoft stock (MSFT) is listed on the Nasdaq exchange. While the company hasn’t tied specific stock price targets to the conference, the announcements signal a fundamental transformation in Microsoft’s product development and artificial intelligence approach.
The Project Solara initiative encompasses a series of experimental devices spanning multiple form factors, from smart speaker-sized units to badge-style devices. Utilizing processors from Qualcomm and MediaTek, these products bypass conventional operating systems completely, instead operating exclusively through AI agents.
Nadella positioned this as an opportunity to fundamentally “rewrite the rules” governing platform development, offering developers and business users unprecedented freedom to create agent-centric hardware solutions.
Surface RTX Spark Dev Box Unveiled In the personal computing segment, Microsoft introduced the Surface RTX Spark Dev Box, equipped with Nvidia’s RTX Spark processor. Nadella described it as his “dream machine” and revealed he had personally joined the waiting list.
This system successfully executed a 120-billion parameter AI model entirely on-device—a capability beyond most existing personal computers. Microsoft simultaneously launched a collaborative laptop with Nvidia this week, directly challenging Apple’s dominance in the premium computer segment.
Industry analysts suggested that widespread enterprise adoption of these advanced systems may require considerable time.
Microsoft additionally announced efforts to adapt OpenClaw—open-source technology for coordinating multiple AI agents—for secure enterprise deployment on Windows platforms. This software has already contributed to increased Mac computer sales for Apple in the Chinese market.
MAI Thinking-1 Launch and Direct Anthropic Competition Microsoft’s artificial intelligence division introduced MAI Thinking-1, the company’s inaugural in-house reasoning model, which Microsoft claims achieves parity with Anthropic’s Claude Opus 4.6. Anthropic has subsequently launched Opus 4.8.
This model forms part of Microsoft’s strategic effort to develop cutting-edge AI capabilities independently of OpenAI, despite years of financial backing. A restructured partnership agreement finalized in April granted Suleyman’s division autonomy to pursue proprietary model development.
AI division leader Mustafa Suleyman spoke candidly about competitive dynamics in a Bloomberg interview: “Anthropic is extremely expensive, and I think many people are urgently looking for alternatives.”
He continued: “We pay a lot of money to Anthropic — so our goal is to reduce and ultimately eliminate that cost.”
Microsoft positioned its latest coding model as delivering equivalent performance to Anthropic’s Opus 4.6 while offering superior cost efficiency—identifying pricing as a crucial competitive differentiator.
Appian CEO Matt Calkins provided broader market context: “We are in the era of subsidies for AI. When OpenAI and Anthropic go public, these prices will probably increase substantially.”
Anthropic submitted its IPO prospectus confidentially to the Securities and Exchange Commission this week. OpenAI is anticipated to follow with its own filing in the near term.
Regarding healthcare initiatives, Microsoft revealed a collaboration with Mayo Clinic focused on building frontier healthcare artificial intelligence, merging Microsoft’s computational and reasoning infrastructure with Mayo’s extensive clinical datasets.
