Microsoft finishes pretraining 2.8T-parameter MAI-Base-2
Microsoft AI details MAI-Base-2, a 2.8T-parameter MoE model with 80B active, trained on 50T tokens across 16,000 GB200s over 38 days at 21% MFU.
Large Models News and Analysis
134 verified stories covering Large Models, product updates and industry developments. Page 1 of 2
Anthropic Q3 Revenue Hits $16.8B as Growth Slows
Anthropic reportedly told investors its Q3 revenue reached $16.8 billion, up about 45% quarter on quarter, with adjusted profit for a second straight quarter.
Google's AMIE hits The Lancet: 98 chats, zero halts
Google's conversational diagnostic AI AMIE has its first prospective real-clinic results in The Lancet: 98 completed chats, no safety stops, and a 90% differential hit rate.
Meta Launches Muse Wonder as API Output Price Rises Nearly 5x
Meta's Muse Wonder lifts the Artificial Analysis index to 52, but output tokens now cost $20 per million, up from $4.25. The stronger Full tier is Muse Code only.
Microsoft's Decision-1 Is Built on Qwen 9B
Microsoft's new Decision-1 model for routing and classification is post-trained on Alibaba's open-source Qwen3.5-9B, priced at $0.042 per million input tokens with free output.
OpenAI's run-rate revenue is near $50B, $20B below earlier reports
OpenAI told investors its annualized revenue is close to $50 billion as of late September, well short of the roughly $70 billion that circulated a week earlier.
ChatGPT Switches All Users to GPT-6, Replies Gain Buttons and Forms
OpenAI made GPT-6 the default model for all ChatGPT tiers and launched Intelligent UI, so replies can include buttons, forms and charts, not just text.
OpenAI Publishes 722 AI-Generated Math Manuscripts
OpenAI released 722 AI-generated math manuscripts on GitHub, covering about 4,000 problems with partial Lean-verified proofs.
Mistral Unveils 1-Trillion-Parameter Large 4 Preview
Mistral's 1-trillion-parameter Large 4 preview scores 38 on Artificial Analysis's Intelligence Index, tying GPT-6 Luna, with open weights due by month's end.
Reflection AI debuts 501B open model Beam
Reflection AI unveils Beam, a 501B-parameter open MoE model claiming to rival Chinese models like GLM 5.2 and Kimi K3, with weights due later this month.
Astra, Sol Subscription Speed Jumps to 50 Tokens/Second
OpenAI boosts default output speed for GPT-6 Astra and GPT-6.1 Sol by about 50%, from 30 to 50 tokens per second, covering partner apps linked via ChatGPT sign-in too.
Ai2 Open-Sources 8B Model for Research Reports, 3.5x Faster
Ai2 open-sourced AstaBrief 8B, which writes cited research reports 3.5x faster than a Claude-based pipeline, with similar user approval.
GPT-6.1 Sol Arrives With API Pricing a Fifth of Astra's
OpenAI's new GPT-6.1 Sol matches flagship-level coding and computer-use scores while costing a fifth of GPT-6 Astra's API price, benchmarks show.
Claude Sonnet 5.5 Launches: 30% Faster at No Extra Cost
Anthropic's Claude Sonnet 5.5 claims 30% speed gains and lower costs, but third-party testing shows a different picture at the highest reasoning tier.
Anthropic Engineer: Claude Learned to Write for AI, Not People
Anthropic engineer Jackson Kernion: RL rewards pushed Claude toward dense, jargon-heavy prose. Opus 5.5 targets sentence clarity, no benchmark numbers shared.
vLLM Adds Built-In Text Watermarking With Minimal Throughput Cost
vLLM added a Gumbel-max text watermark toggled by one server flag, with near-zero throughput cost — but code and short replies are hard to detect reliably.
GPT-6 Sol and Luna Launch as OpenAI Halves API Prices
OpenAI's new GPT-6 Sol and Luna cost half as much as GPT-5.6 per API call, while third-party testing shows capability scores holding roughly steady as costs drop sharply.
Claude Opus 5.5 Launches, Cuts Running Costs 40%
Anthropic's Claude Opus 5.5 lowers API pricing and overall running costs by 40% versus Opus 5, with faster output and a new fast mode.
StepFun Launches Step 5 Preview, Claims Cost 1/8 of Opus 5
StepFun's new Step 5 Preview MoE model claims a top-three open-weight benchmark score and a cost one-eighth of Claude Opus 5, with full weights due October 15.
Gemini Distillation Service Enters Preview, Locked to One Region
Google Cloud's Gemini distillation service enters preview, but the teacher model, student model, and region are all fixed for now.
DeepSeek's New Model Matches Astra on Coding at 1/15 the Cost
New Fireworks AI benchmark data shows DeepSeek-V4.1-Flash matching GPT-6 Astra's score on agentic coding tasks while costing 15 times less per task.
Siri AI Launches in Beta, English Only at First
Apple's overhauled Siri AI arrives with iOS 27 in an English-only beta, with the EU, China and a daily usage cap all left out of this first wave.
DeepSeek Ships V4.1 Flash, Reroutes Pro Traffic Monday
DeepSeek launched V4.1 Flash on Sept 10 with new pricing, and starting Sept 14 all deepseek-v4-pro requests reroute to the smaller model at its rates.
Data Drives 3x More Pretraining Progress Than Model Design
A controlled study finds pretraining data gains outpaced model design gains 12x vs 3.7x from 2019-2025, though annualized rates trail industry estimates.
OpenAI publishes Navier-Stokes proof, peer review still pending
OpenAI says agent swarms produced a Navier-Stokes proof in 88 hours, but no mathematician has verified it and a rival team disputes how the work was shared.
DeepSeek Releases V4.1 Flash Preview, Expiring September 10
DeepSeek opened a V4.1 Flash preview with native multimodal support, but the model ID expires September 10 — under 48 hours after launch.
Mathematician Goes Public on Disputed OpenAI Proof Call
NYU mathematicians published new blowup proofs and a statement disputing OpenAI's account of a call over credit for a model-assisted proof.
OpenAI Chief Scientist Says Chain-of-Thought Monitoring Is Weakening
OpenAI chief scientist Jakub Pachocki says chain-of-thought monitoring is weakening, and no lab has solved alignment enough to scale responsibly at full speed.
Moonshot AI Seeks 30% Revenue Cut for Kimi K3
Moonshot AI is negotiating a 30% cloud revenue share for open-weight Kimi K3 with Microsoft, Amazon and Google, testing a first-of-its-kind deal that would beat the industry's old 20% ceiling.
Muse Spark 1.3 Hits 75.4 on Coding Benchmark, Tops GPT-5.6 Sol
Meta's Muse Spark 1.3 scores 75.4 on DeepSWE, beating GPT-5.6 Sol and Claude Opus 5 in coding, but lags all three agentic benchmarks it was tested on.
LLM Token Price Index Drops Below $1, Half Its Summer Peak
Silicon Data's LLM token price index closed below $1 for the first time, down more than half from its summer peak, squeezing frontier labs' margins ahead of IPO season.
Qwen3.8-Max Takes Coding Crown at Quarter the Price of Rivals
Alibaba's Qwen3.8-Max tops Code Arena's frontend coding leaderboard with a score of 1691, priced at roughly a quarter of the runner-up's rate.
Claude Fable 5.1 Cuts Cache-Read Price to a Quarter
Fable 5.1 keeps input and output pricing flat, cuts cache-read cost to a quarter for long agent sessions, but breaking changes shift real cost to output tokens.
South Korea's Free AI for All Rides on 512 B200 GPUs
South Korea picked SK Telecom, KT, and Kakao consortia to run a free, token-unlimited AI service for all citizens using just 512 Nvidia B200 GPUs.
Google Employees Get Early Access to Gemini 3.8 Flash
Business Insider reports Google employees are testing a Gemini 3.8 Flash preview via internal platform Jetski, as release cycles compress toward monthly.
Tencent Opens Hunyuan Hy4, Narrowly Tops GLM and Kimi in Blind Test
Tencent open-sources Hunyuan Hy4 preview, a 770B MoE model with 6.4% activation and 1M-token context, narrowly beating GLM-5.3 and Kimi K3 in blind tests.
China's Daily AI Token Calls Surge to 500 Trillion in Three Months
China's daily AI token calls hit 500 trillion in June, up from 140 trillion in March, as AI agents push inference costs past training costs.
Google's New Transcription Model Cuts Word Error Rate to 2.6%
Google's Gemini 3.5 Transcribe cuts word error rate to 2.6%, speeds up transcription 70% over Chirp 3, and supports 85+ languages with real-time and pre-recorded modes.
OpenAI brings back the five-hour cap for ChatGPT Plus
OpenAI is reinstating the five-hour usage limit for Plus accounts on ChatGPT Work and Codex, while a routing bug quietly downgraded 3% of Pro and Thinking requests to GPT-5.5-mini.
Qwen3.8-Flash-Next Launches Tonight, Architecture Jumps Ahead to Qwen4
Alibaba's Qwen3.8-Flash-Next launches tonight on the unreleased Qwen4 architecture, a multimodal MoE with under 5% parameter activation aimed at cutting inference cost.
Anonymous 'Ox Alpha' Model Offers Free 1M-Token Context
An unnamed model called Ox Alpha appeared on OpenRouter with a free 1.05-million-token context window, and fingerprint clues point to Zhipu AI's GLM lineup.
Time for Open Models to Catch Up Is Halving Each Generation
SemiAnalysis charts three years of open-vs-closed AI gaps and finds the catch-up window keeps halving, with Kimi K2.6 closing on Opus 4.5 in just 4.8 months.
Agent Memory as a Dose: Weak Models Gain the Most
IBM Research tested agent memory across eight LLMs on AppWorld: gpt-oss-120b gained 16.1 points with selective retrieval, while GLM-5 saw zero change.
OpenRouter Now Bills AI Agents Individually
OpenRouter's new Activity dashboard and beta Analytics API break down AI spending by agent, model, and single request, giving teams a traceable view of agent costs.
Google puts sign-language AI into Pixel
Google DeepMind's SL2T brings ASL-to-English input to Gboard and Live Transcribe on Pixel 11, with a privacy design based on pose landmarks rather than raw video.
DeepSeek Pro adds 384K output
DeepSeek's V4 Pro API now points to the 0813 version, pairing a 1M-token context and 384K output ceiling with a clear price warning.
Tencent AI capex jumps to RMB52.8B
Tencent's second-quarter report shows AI moving from product demos into capex, cash flow and WeChat-scale deployment tests.
ModelBest starts A-share IPO tutoring
ModelBest has entered IPO tutoring with CITIC Securities, pushing China's edge-model race from open-source downloads into audited revenue, governance and deployment tests.
Former Qwen lead bets on agents
Junyang Lin, former technical lead of Alibaba Qwen, has launched Pragmatik Labs in Shanghai with backing from Gaorong, HSG, Tencent and Shanghai Future Industry Fund, aiming at digital and physical agents.
BigBang Lets Models Write Tasks
BigBang-V1 turns synthetic-data production into an agentic loop: models propose, solve, verify, and filter frontier tasks.
Palantir Turns Earnings Into an AI Sovereignty Pitch
Palantir's Q2 results turned a fast-growth earnings report into a sharper fight over enterprise AI sovereignty, data control and the role of model labs.
Inkling Small Closes In on 975B Flagship
Thinking Machines released Inkling-Small with open weights, 276B total parameters, 12B active parameters, multimodal input and benchmark scores close to the 975B Inkling.
DeepSeek Opens the Final V4-Flash API
DeepSeek put V4-Flash 0731 into API beta, added Responses API support for Codex, and reported large agent-benchmark gains from post-training alone.
Kimi Valuation Hits $35B
Moonshot AI has reportedly closed more than $3.5 billion in F-round financing at a $35 billion post-money valuation after opening Kimi K3 weights.
Tencent Open-Sources an Inference Speedup Stack
Tencent Hunyuan released AngelSpec, an open-source framework for training speculative-decoding drafters and reducing large-model inference cost.
Kimi K3 Falls Behind in Cyber Tests
UK AISI and US CAISI put Moonshot AI's Kimi K3 through cyber benchmarks, finding strong general-model momentum but a clear gap on exploit chains.
Poolside Releases 118B Open Coding Model
Laguna S 2.1 gives enterprises a self-hosted coding-agent base, trading frontier scores for local control and clearer deployment economics.
Gemini 3.5 Pro slips by months
Google is reportedly months behind on Gemini 3.5 Pro as coding performance becomes the main test of its flagship model.
Kimi K3 arrives with 2.8T parameters
Moonshot AI has launched Kimi K3 with 2.8 trillion parameters, a 1M-token context window and published API pricing, while the open-weight proof still depends on weights and model cards.
Agnes-2.5-Flash uses a free coding model to pressure agentic developer tools
Agnes-2.5-Flash uses a free coding model to pressure agentic developer tools. The piece reviews the verified facts and why the signal matters beyond one announcement.
MiniMax raises another HK$16B
MiniMax raises another HK$16B. A concise localization of the verified facts and the industry signal behind the story.
PrismML squeezes a 27B model onto an iPhone
PrismML squeezes a 27B model onto an iPhone. A concise localization of the verified facts and the industry signal behind the story.
Yuanli trains DM0.5 on 150,000 hours of robot data
Yuanli trains DM0.5 on 150,000 hours of robot data is reframed for global technology readers, with the key numbers, product claims and open questions kept intact.
01.AI launches three decision AI products
01.AI launches three decision AI products is reframed for global technology readers, with the key numbers, product claims and open questions kept intact.
Oxford study says AI rewriting can steer opinions
Oxford study says AI rewriting can steer opinions is reframed for global technology readers, preserving the key numbers, claims and open questions.
Meta says Watermelon is catching up with GPT-5.5
Internal AI benchmarks, compute scale and Meta’s model strategy shift.
DeepSeek Adds Peak-Hour Pricing for V4 API
DeepSeek plans a mid-July V4 release and will double API prices during two China workday peak windows to smooth demand.
Gemini 3.5 Pro Slips From June to July
Google moved Gemini 3.5 Pro from a June release into July, leaving the flagship model limited to a small Vertex AI preview.
GPT-5.6 Arrives Behind a Government Review Gate
OpenAI introduced GPT-5.6 Sol, Terra and Luna with lower token prices, but access is initially limited to about 20 government-reviewed partners.
Meta Restricts Claude Code and Codex for AI Teams
Meta told some applied-AI engineers to seek approval before using Claude Code or Codex, citing model distillation and contract risks.
OpenAI Prices GPT-5.6 in Three Tiers
OpenAI split GPT-5.6 into Sol, Terra and Luna, using Sol for top performance, Terra for mainstream work and Luna for cheaper high-volume use.
DeepSeek opens DSpark for faster inference
DeepSeek and Peking University open sourced DSpark, an inference framework that raises generation speed by 60% to 85% and can lift single-GPU throughput as much as 6.6 times.
Anthropic frames Claude Science as an auditable workbench
Claude Science is a research workbench rather than a new model, designed to collect scientific data and make AI-assisted results auditable and rerunnable.
Claude Sonnet 5 targets cheaper agents
Anthropic positioned Claude Sonnet 5 as a mid-tier model for autonomous agent work, with promotional pricing below the Opus tier.
Gemini 3.5 Pro slips past June as prediction markets price July
Gemini 3.5 Pro slips past June as prediction markets price July
Google forms a coding push as AI-written code stalls near half
Google forms a coding push as AI-written code stalls near half
Lindy swaps Claude for DeepSeek and says costs fell by 90%
Lindy swaps Claude for DeepSeek and says costs fell by 90%
South Korea puts $518B behind four AI chip clusters
South Korea puts $518B behind four AI chip clusters
Wix’s Base44 ships its own model as ARR reaches $150M
Wix’s Base44 ships its own model as ARR reaches $150M
DeepSeek brings peak-and-off-peak pricing to model APIs
DeepSeek brings peak-and-off-peak pricing to model APIs
GPT-5.4 lifts Chan-Lam yields in wet-lab tests
OpenAI and Molecule.one used GPT-5.4 to identify TEMPO as an additive, raising average Chan-Lam coupling yields from 16.6% to 25.2% across high-throughput experiments.
GPT-5.6 leaks point to a 1.5M-token window
Unconfirmed OpenAI signals suggest GPT-5.6 may focus on longer context, slower but stronger reasoning, and agentic coding.
Altman says AI skeptics underestimated scaling
Sam Altman frames the scaling debate around new reasoning results, while conceding that long-horizon work remains a hard gap.
OpenAI replays real chats to test new models
OpenAI’s Deployment Simulation uses 1.3 million consented historical conversations to predict how unreleased models will behave in production, including agentic tool-use failures.
Kimi sends 300 agents to forecast the 2026 World Cup
Moonshot AI used a cluster of Kimi agents to simulate all 104 matches and identify probability gaps, including a higher chance for Germany than betting markets imply.
Chinese models reset the floor for AI pricing
Enterprise workload estimates show Claude, ChatGPT, DeepSeek, Kimi and GLM separated by nearly 9x in cost, forcing OpenAI and Anthropic toward price pressure.
Nadella says token use has to earn its cost
Microsoft's CEO called himself a token-maxer but warned that each extra token should be matched by real productivity gains, not just bigger AI bills.
AI memory can make models agree at the cost of accuracy
AI memory can make models agree at the cost of accuracy.
A Stroop-style test trips up leading language models
A new attention test pushed major models into sharp accuracy drops, reminding buyers that reasoning claims still need careful benchmarks.
Google opens DiffusionGemma for faster text generation
DiffusionGemma applies diffusion-style decoding to text, generating chunks in parallel and showing another path beyond autoregressive language models.
Cohere releases North Mini Code for local coding
Cohere’s 30B-parameter MoE model activates about 3B parameters, supports 256K context and is released under Apache 2.0.
Anthropic prices Fable 5 at twice Opus
Anthropic opened Claude Fable 5 broadly at $10 per million input tokens and $50 per million output tokens, double Opus 4.8.
Anthropic releases Claude Fable 5 with safety fallbacks
Claude Fable 5 is Anthropic’s strongest public model, adapted from Mythos and designed to fall back to Opus 4.8 for cyber, bio, chemical and distillation risks.
Anthropic opens Claude Fable 5 to the public
Anthropic has made Claude Fable 5 broadly available, using classifiers to route sensitive cyber, bio, chemical and distillation requests away from the strongest model while keeping most conversations on Fable.
Xiaomi MiMo claims 1,000 tokens per second on standard GPUs
Xiaomi MiMo claims 1,000 tokens per second on standard GPUs. The story explains the announcement, the strategic context and the practical risks for the people or companies affected.
Microsoft unveils seven in-house MAI models
At Build 2026, Microsoft showed a full in-house model stack from reasoning to image, voice, transcription and coding, signaling a cheaper fallback beside OpenAI.
Microsoft shows seven MAI models to reduce OpenAI reliance
At Build, Microsoft presented seven in-house models, including MAI-Thinking-1, a 35B active-parameter MoE reasoning model trained from scratch without distillation.
Trump Order Gives US Cyber Defenders a 30-Day Look at Frontier AI Models
The executive order keeps model review voluntary but asks companies to give government cyber teams early access before wider release.
Alibaba pitches Qwen3.7-Plus as a self-checking agent model
Alibaba released Qwen3.7-Plus on June 2, highlighting multimodal input, tool use, code generation and autonomous testing as its next agent bet.
Google quietly offers to buy private app code for AI training
A reported Google Play pilot asks developers to sell source code, including dormant projects, while the email itself avoids saying AI.