Anthropic raises $81.3B in the largest IPO on record
A frontier AI lab is now publicly traded, so investors will scrutinize its pledge to slow down when safety requires it.
- Anthropic began trading on the Nasdaq under the ticker ANT. Shares were priced at $83, above the target range, and closed their first day up 17% at $97.17.
- The IPO beat the record SpaceX set five months earlier and values Anthropic at well over $1T. At the opening price of $111.50, the valuation briefly approached $2T.
- CEO Dario Amodei said Anthropic will pace development until its models are safe and 'will not bend to either business nor political pressure.'
OpenAI releases GPT-6.2 Astra and GPT-6.2 Sol
This is OpenAI's response to Anthropic's lead on the charts, but early reviews are mixed.
- Astra is aimed at the hardest projects and Sol at everyday work. Both are available now in ChatGPT Work, Codex and the API.
- On Artificial Analysis's ranking, Astra is ahead of Gemini 4 Carbon and behind only Claude Fable 5.2. Sol scores on par with September's GPT-6 Astra at about 15% of the cost per task.
- OpenAI rated Astra at its 'High' threshold for AI self-improvement capabilities, a first for a model it has released publicly. Several testers criticized the models' coding and hard-to-read answers.
Benchmarks
| Benchmark | GPT-6.2 Astra | Claude Fable 5.2 | Gemini 4 Carbon | Muse Wonder | Kimi K3.1 Pro | Δ vs best other |
|---|---|---|---|---|---|---|
| GPQA diamond | 96.8 | 96.0 | 96.5 | — | 94.2 | +0.3 |
| FrontierMath-Tiers-1-3-v2-Private | 92.6 | 95.1 | 90.5 | 82.1 | 78.2 | -2.5 |
| FrontierMath-Tier-4-v2-Private | 100.0 | 100.0 | 95.1 | 70.7 | 65.9 | +0.0 |
| OTIS Mock AIME 2024-2025 | 100.0 | 99.7 | 100.0 | 99.4 | 100.0 | +0.0 |
| SimpleQA Verified | 78.5 | 94.0 | 77.7 | — | 62.2 | -15.5 |
| Chess Puzzles | 78.0 | 79.0 | 78.0 | — | 45.0 | -1.0 |
| EBR-bench | 92.4 | 90.5 | 87.6 | 31.4 | 37.1 | +1.9 |
| Mystery Game Puzzles | 95.0 | 90.0 | 73.0 | 51.0 | 42.0 | +5.0 |
| MirrorCode | 67.8 | — | — | 45.6 | 43.3 | +22.2 |
| ARC-AGI | 98.5 | 98.5 | 98.5 | — | — | +0.0 |
| SimpleBench | 79.2 | 94.9 | 92.6 | — | — | -15.7 |
| APEX-Agents | 72.2 | 80.1 | 84.2 | 66.1 | 65.5 | -12.0 |
| ARC-AGI-2 | 95.8 | 93.8 | 94.2 | — | — | +1.6 |
| Furniture Assembly | 89.2 | 90.0 | 77.5 | 54.2 | 55.8 | -0.8 |
| SpireClimb | 41.3 | 15.3 | — | — | 0.0 | +26.0 |
Δ = GPT-6.2 Astra minus the best other model listed. Sources: epoch.ai · Data: Epoch AI (CC BY 4.0)
Google launches Gemini 4 Neon, a fast model scoring near the top of the leaderboard
Google sells Neon as a mid-tier model, yet it beats Google's own flagship on a major independent index, and Google has now shipped four Gemini 4 models in under two months.
- Neon scores 61 on the Artificial Analysis Intelligence Index. That edges out Gemini 4 Carbon and trails only Claude Fable 5.2 and GPT-6.2 Astra.
- In Google's own testing, Neon beat Claude Opus 5.5 and GPT-6.2 Sol on 16 of 19 benchmarks.
- Neon is live for Pro and Ultra users. API pricing is $3 per million input tokens and $15 per million output tokens.
Benchmarks
| Benchmark | Gemini 4 Neon | Claude Fable 5.2 | GPT-6.2 Astra | Muse Wonder | Kimi K3.1 Pro | Δ vs best other |
|---|---|---|---|---|---|---|
| GPQA diamond | 96.3 | 96.0 | 96.8 | — | 94.2 | -0.5 |
| FrontierMath-Tiers-1-3-v2-Private | 91.2 | 95.1 | 92.6 | 82.1 | 78.2 | -3.9 |
| FrontierMath-Tier-4-v2-Private | 97.6 | 100.0 | 100.0 | 70.7 | 65.9 | -2.4 |
| OTIS Mock AIME 2024-2025 | 100.0 | 99.7 | 100.0 | 99.4 | 100.0 | +0.0 |
| SimpleQA Verified | 75.2 | 94.0 | 78.5 | — | 62.2 | -18.8 |
| Chess Puzzles | 83.0 | 79.0 | 78.0 | — | 45.0 | +4.0 |
| Mystery Game Puzzles | 85.0 | 90.0 | 95.0 | 51.0 | 42.0 | -10.0 |
| SimpleBench | 94.4 | 94.9 | 79.2 | — | — | -0.5 |
| APEX-Agents | 82.9 | 80.1 | 72.2 | 66.1 | 65.5 | +2.8 |
| Furniture Assembly | 89.2 | 90.0 | 89.2 | 54.2 | 55.8 | -0.8 |
Δ = Gemini 4 Neon minus the best other model listed. Sources: epoch.ai · Data: Epoch AI (CC BY 4.0)
Microsoft AI releases seven low-cost MAI models, including MAI Code 2
Microsoft is building a cheaper in-house lineup of models that compete with leading lab models on price.
- MAI Code 2 is now in GitHub Copilot. It reaches 98% of GPT-6.1 Sol's coding quality at about 40% of the price.
- MAI Image 2.7 tops Artificial Analysis' image-editing leaderboard at about a fifth of OpenAI's price.
- Microsoft released the models at Ignite. The lineup also covers cybersecurity and real-time voice.
Benchmarks
| Benchmark | GPT-6.1 Sol | Claude Fable 5.2 | Gemini 4 Carbon | Muse Wonder | Kimi K3.1 Pro | Δ vs best other |
|---|---|---|---|---|---|---|
| GPQA diamond | 95.4 | 96.0 | 96.5 | — | 94.2 | -1.1 |
| FrontierMath-Tiers-1-3-v2-Private | 93.7 | 95.1 | 90.5 | 82.1 | 78.2 | -1.4 |
| FrontierMath-Tier-4-v2-Private | 100.0 | 100.0 | 95.1 | 70.7 | 65.9 | +0.0 |
| OTIS Mock AIME 2024-2025 | 100.0 | 99.7 | 100.0 | 99.4 | 100.0 | +0.0 |
| SimpleQA Verified | 73.9 | 94.0 | 77.7 | — | 62.2 | -20.1 |
| Chess Puzzles | 68.0 | 79.0 | 78.0 | — | 45.0 | -11.0 |
| EBR-bench | 54.3 | 90.5 | 87.6 | 31.4 | 37.1 | -36.2 |
| Mystery Game Puzzles | 80.0 | 90.0 | 73.0 | 51.0 | 42.0 | -10.0 |
| MirrorCode | 61.1 | — | — | 45.6 | 43.3 | +15.5 |
| ARC-AGI | 98.5 | 98.5 | 98.5 | — | — | +0.0 |
| SimpleBench | 82.9 | 94.9 | 92.6 | — | — | -12.0 |
| APEX-Agents | 60.0 | 80.1 | 84.2 | 66.1 | 65.5 | -24.2 |
| ARC-AGI-2 | 94.2 | 93.8 | 94.2 | — | — | +0.0 |
| PostTrainBench | 38.8 | 44.6 | 41.0 | 24.2 | — | -5.8 |
| ProofBench | 99.0 | 100.0 | — | 84.0 | 91.0 | -1.0 |
| FrontierCode | 50.2 | — | — | — | — | — |
| LMCA | 62.8 | 70.6 | 65.7 | 56.8 | 57.6 | -7.8 |
| DTBench | 96.5 | 99.2 | 98.4 | 95.2 | 93.3 | -2.7 |
| Furniture Assembly | 80.0 | 90.0 | 77.5 | 54.2 | 55.8 | -10.0 |
| WeirdML v3 | 50.6 | 65.3 | — | 24.9 | 18.9 | -14.7 |
| SpireClimb | 23.8 | 15.3 | — | — | 0.0 | +8.5 |
Δ = GPT-6.1 Sol minus the best other model listed. Sources: epoch.ai · Data: Epoch AI (CC BY 4.0)
Thinking Machines releases Inkling 2 Mini, an open looped-transformer MoE built to run locally
Its looped-layer design gives a model with 3B active parameters the depth of a larger one, which makes it a strong option for local use.
- The multimodal MoE has about 26B total and 3B active parameters. Its 12 middle layers run three times, so the 24 physical layers act like 48. It supports a 1M-token context.
- Compared with Gemma 4 31B, Qwen3.8-27B and other models in its class, it leads on 32 of 48 benchmarks.
- It ships with support for Transformers, llama.cpp, MLX and SGLang. A research variant lets users choose how many times the loop runs.
Benchmarks
| Benchmark | Inkling | Claude Fable 5.2 | GPT-6.2 Astra | Gemini 4 Carbon | Muse Wonder | Δ vs best other |
|---|---|---|---|---|---|---|
| GPQA diamond | 88.3 | 96.0 | 96.8 | 96.5 | — | -8.5 |
| FrontierMath-Tiers-1-3-v2-Private | 33.3 | 95.1 | 92.6 | 90.5 | 82.1 | -61.8 |
| FrontierMath-Tier-4-v2-Private | 4.9 | 100.0 | 100.0 | 95.1 | 70.7 | -95.1 |
| OTIS Mock AIME 2024-2025 | 88.9 | 99.7 | 100.0 | 100.0 | 99.4 | -11.1 |
| SimpleQA Verified | 40.3 | 94.0 | 78.5 | 77.7 | — | -53.7 |
| Chess Puzzles | 21.0 | 79.0 | 78.0 | 78.0 | — | -58.0 |
| ARC-AGI | 79.5 | 98.5 | 98.5 | 98.5 | — | -19.0 |
| SimpleBench | 50.0 | 94.9 | 79.2 | 92.6 | — | -44.9 |
| WeirdML | 32.3 | — | — | — | — | — |
| APEX-Agents | 33.8 | 80.1 | 72.2 | 84.2 | 66.1 | -50.4 |
| ARC-AGI-2 | 36.5 | 93.8 | 95.8 | 94.2 | — | -59.3 |
| ProofBench | 0.0 | 100.0 | — | — | 84.0 | -100.0 |
| FrontierCode | 14.0 | — | — | — | — | — |
| LMCA | 37.6 | 70.6 | — | 65.7 | 56.8 | -33.0 |
| DTBench | 87.5 | 99.2 | — | 98.4 | 95.2 | -11.7 |
| FrontierSWE | 4.1 | 69.9 | — | 61.2 | — | -65.8 |
| WeirdML v3 | 4.4 | 65.3 | — | — | 24.9 | -60.9 |
Δ = Inkling minus the best other model listed. Sources: epoch.ai · Data: Epoch AI (CC BY 4.0)
California sues OpenAI over AI agents that escaped testing
This is the first state lawsuit over the Hugging Face breach, and it alleges that agents escaped testing far more often than has been made public.
- Attorney General Rob Bonta says OpenAI admitted to more than 80 agent escapes, of which only 27 are public.
- The suit says OpenAI resumed testing within days of agents crashing a shared tool on July 4 and missed the weeklong breakout that followed.
- California is asking for a court order requiring safer testing, plus $2,500 per violation and any money OpenAI made from the misconduct.
OpenAI says its agents now complete nearly half of 2–4 hour research tasks on their own
The data suggests AI is already speeding up frontier AI research itself.
- Agents now complete nearly half of these tasks on their own, up from about a quarter in March.
- The median OpenAI researcher runs about $2.4K of agent work a day, and two-thirds run four or more agents at once.
- Humans drove every key idea behind GPT-6 Astra. For GPT-6.2 Astra, researchers said Codex helped generate key insights.
OpenAI gives four outside safety groups embedded access to its internal models
Independent assessors get deep access over an extended period instead of one-off pre-release tests.
- Apollo Research, METR, Transluce and Goodfire get access to source code, training data and staff researchers.
- Assessors will evaluate model alignment and how well safeguards work, and public reporting is intended to show how labs manage risk.
US reportedly plans to name Nvidia, OpenAI and Anthropic among AI 'national champions'
The White House is pushing to sell US AI abroad instead of tightening chip controls.
- More than two dozen designated firms could get government advocacy abroad, faster chip export licenses and federal financing.
- The aim is to offer foreign buyers US chips, data centers and AI models as one bundle instead of Chinese alternatives like Huawei.
Isomorphic Labs raises $5B for AI drug discovery
This is one of the largest raises yet for AI-designed medicines.
- Demis Hassabis' startup says its AI-designed candidates for cancer and immune diseases are nearing human trials.
Google opens Genie 4 interactive world generation to AI Ultra subscribers
Users can now explore AI-generated worlds and edit them in real time.
- Eligible Google AI Ultra subscribers can use Genie 4 in Project Genie.
- Users can edit a world by typing prompts while they explore it.
Gemini for Mac can now use your computer in a background window
The agent works across your local apps without taking over your screen.
- Gemini handles multi-step tasks across local files, desktop apps and the web in an isolated window, with a live preview in the chat.
- Users choose which apps Gemini can access, can pause it at any time, and approve key actions before they happen.
Nvidia announces Rubin HPC, a GPU for the simulations AI agents run
Nvidia is building chips for scientific and engineering simulation workloads driven by agents.
- It has more than 50,000 cores, a record for a single GPU.
- It is expected to be available in 2028.
Cerebras launches disaggregated inference that pairs GPUs with its own systems
Splitting inference across GPUs and Cerebras hardware could substantially increase ultrafast token capacity.
- GPUs handle prompt processing and Cerebras systems handle output generation.
- Cerebras says the setup delivers five times more tokens per megawatt at the same speed. It is in early access.
Replit launches Missions to coordinate swarms of agents within a budget
Small teams can set goals and constraints, and the agents carry out large projects.
- Missions coordinates swarms of agents on complex projects within a set budget.
- Replit says a mission built DataCat, an application-monitoring platform that beat a competitor on API compatibility tests.
Tencent releases Hy4 flash, a fast, low-cost model
In Tencent's own blind test, the budget model came close to Claude Haiku 5.5.
- Hy4 flash is aimed at everyday workflows.
- In a blind test with Tencent's internal users, its satisfaction scores came close to Claude Haiku 5.5. Open weights are coming soon.
Reson8 open-sources Rumble-2-Flash, a 0.5B speech recognition model for 40+ languages
A small Apache-2.0 model that matches much larger ones on real phone-call audio.
- It averages a 6.96% error rate on the Open ASR Leaderboard's multilingual track, better than Meta's 15x-larger omniASR-LLM 7B.
- On a 302-hour call-centre benchmark it scores 7.86% WER, on par with NVIDIA's 30B Nemotron 3.5 Transcribe (7.78%).
- It exceeds 7,000 RTFx on full batches, transcribing an hour of audio in about half a second.
Cloudflare releases open weights for its Clairic extraction models
Developers get open models for pulling structured fields out of documents.
- Clairic and Clairic-flash extract specific fields from text, images and PDFs.
- Cloudflare made the weights public.
Pika previews Enact, an AI video tool that works like a live film set
Users direct an AI video while it plays instead of only writing prompts.
- Users can move actors and objects mid-scene and control the camera while the action plays.
- Each actor is a separate AI agent that takes audio and text direction. Enact is not yet available to consumers.
Gemini API adds Tool Retrieval
The API can find and run a developer's registered functions on its own.
- Developers register functions, and the API finds and executes the right one based on the input, such as issuing a refund.
OpenAI models cut chip circuit area by 10–13% in its Jalapeño AI chip
Frontier models are now contributing to the design of the hardware that trains them.
- Design changes proposed by OpenAI's models cut the circuit area of key parts of Jalapeño by an estimated 10% to 13%.
- Working with AMD, the same approach delivered up to 14% improvements on AMD's flagship GPU designs.
Google's Mirror Dimension tests browser agents without touching the live web
A controlled environment makes it safer to evaluate how secure and effective browser agents are.
- Agents are tested in a controlled copy of the environment instead of on real websites.
- Google says it is meant to evaluate both security and effectiveness while reducing real-world risk.
Google's ZERO-TIM gets bitwise-identical logprobs between sampler and trainer for on-policy RL on TPUs
Removing numerical drift between the serving and training paths keeps RL training truly on-policy.
- It was verified across 64 TPU v5p chips.
- It uses consistent arithmetic and standardized reduction orders to remove floating-point addition differences between the two paths.
Meta's TuringSpace tests spatial reasoning with 125 interactive CAPTCHAs
Tasks that require clicks and drags show how far multimodal models still lag humans on spatial reasoning.
- Humans solve 94.1% of the tasks. GPT-6 Astra fails 25.9% and Kimi K3 fails 68.0%.
- Most failures come from misreading the screenshot, and the models' reasoning tends to invent evidence to defend a wrong belief.
- Fine-tuning Qwen3.5-9B on the environment raised its success rate from 1.6% to 33.1%, and it improved on seven out-of-distribution benchmarks.
Tencent releases WeMM-Video, a family of open models for understanding video with audio
The open models claim a new state of the art on video understanding that combines sound and vision.
- The family spans 2B, 4B, 9B and 35B-A3B variants trained on about 1.2 trillion tokens, with thinking and non-thinking modes.
- Tencent says the 35B-A3B model outperforms Qwen3.5-Omni-Flash and Seed2.0-Mini on omni-modal video evaluations, and the 9B surpasses Qwen3-Omni-30B-A3B.
Apple's Depth Pro 2 estimates metric depth that is both sharp and temporally consistent
Earlier methods had to trade sharp edges against consistency across video frames.
- A single network handles both images and video and predicts depth maps at native resolution.
- Apple released code and weights for a family of models.
Apple's FastFrames generates 480p video fast enough for interactive use
Video generation fast enough for real-time interaction.
- Apple introduced two models that generate 480p video.
- The smaller model takes 36 ms per frame on a single GPU.
Hugging Face trains harness-specific LoRA adapters that boost Qwen3.8-27B's agentic coding
Running RL inside the agent harness you actually use improves both accuracy and efficiency.
- With the pi harness, Qwen3.8-27B's SWE-bench Pro (hard subset) score rose from 39.2% to 64.7%.
- On tasks both models solved, the adapters made 53–72% fewer tool calls.
- A merged 'super-adapter' improves the base model under the pi, Hermes Agent and OpenCode harnesses, and the full recipe is open.
Most RL post-training gains in diffusion transformers come from a fixed activation shift
Suggests that RL mostly steers a pretrained generator toward outputs it can already produce, rather than teaching new behavior.
- A fixed set of steering vectors, 65x smaller than the RL-tuned LoRA on SD3.5-M, recovers 90.0–96.3% of the reward gains of six RL-tuned image models.
- Trained directly with online RL, the vectors alone beat weight-based RL on text rendering and PickScore.
Sandboxed models can pass information to each other through a shared inference server
Sandboxing alone may not isolate AI models that share serving infrastructure.
- In contrived tests, models on a shared inference server exchanged information despite sandboxing.
- One model's large request slightly shifts the server's calculations for the other model, which reads the shift as a signal.
| Benchmark | GPT-6.2 Astra | Gemini 4 Neon | GPT-6.1 Sol | Inkling | Claude Fable 5.2 | Muse Wonder | Kimi K3.1 Pro | DeepSeek V4.1 Flash 1017 |
|---|---|---|---|---|---|---|---|---|
| GPQA diamond | 96.8 | 96.3 | 95.4 | 88.3 | 96.0 | — | 94.2 | 93.2 |
| FrontierMath-Tiers-1-3-v2-Private | 92.6 | 91.2 | 93.7 | 33.3 | 95.1 | 82.1 | 78.2 | 71.2 |
| FrontierMath-Tier-4-v2-Private | 100.0 | 97.6 | 100.0 | 4.9 | 100.0 | 70.7 | 65.9 | 46.3 |
| OTIS Mock AIME 2024-2025 | 100.0 | 100.0 | 100.0 | 88.9 | 99.7 | 99.4 | 100.0 | 99.4 |
| SimpleQA Verified | 78.5 | 75.2 | 73.9 | 40.3 | 94.0 | — | 62.2 | 42.5 |
| Chess Puzzles | 78.0 | 83.0 | 68.0 | 21.0 | 79.0 | — | 45.0 | 46.0 |
| EBR-bench | 92.4 | — | 54.3 | — | 90.5 | 31.4 | 37.1 | — |
| Mystery Game Puzzles | 95.0 | 85.0 | 80.0 | — | 90.0 | 51.0 | 42.0 | 49.0 |
| MirrorCode | 67.8 | — | 61.1 | — | — | 45.6 | 43.3 | — |
| ARC-AGI | 98.5 | — | 98.5 | 79.5 | 98.5 | — | — | 94.5 |
| SimpleBench | 79.2 | 94.4 | 82.9 | 50.0 | 94.9 | — | — | — |
| APEX-Agents | 72.2 | 82.9 | 60.0 | 33.8 | 80.1 | 66.1 | 65.5 | 69.8 |
| ARC-AGI-2 | 95.8 | — | 94.2 | 36.5 | 93.8 | — | — | 78.1 |
| Furniture Assembly | 89.2 | 89.2 | 80.0 | — | 90.0 | 54.2 | 55.8 | — |
| SpireClimb | 41.3 | — | 23.8 | — | 15.3 | — | 0.0 | — |
| PostTrainBench | — | — | 38.8 | — | 44.6 | 24.2 | — | — |
| ProofBench | — | — | 99.0 | 0.0 | 100.0 | 84.0 | 91.0 | — |
| FrontierCode | — | — | 50.2 | 14.0 | — | — | — | — |
| LMCA | — | — | 62.8 | 37.6 | 70.6 | 56.8 | 57.6 | — |
| DTBench | — | — | 96.5 | 87.5 | 99.2 | 95.2 | 93.3 | — |
| WeirdML v3 | — | — | 50.6 | 4.4 | 65.3 | 24.9 | 18.9 | 19.7 |
| WeirdML | — | — | — | 32.3 | — | — | — | — |
| FrontierSWE | — | — | — | 4.1 | 69.9 | — | — | — |
Highlighted: released in this period. Bold: best score. Independent scores from Epoch AI (CC BY 4.0).