METR estimates Anthropic's AI progress has roughly doubled, mostly driven by its own AI
An outside evaluator with 11 weeks of inside access says the AI-building-AI loop is already running at a frontier lab.
- METR's report, based on 11 weeks of access to Anthropic's internal data and staff interviews, estimates a 1.3x to 3x speedup over the company's 2024-25 pace, crediting AI for about 80% of the faster progress.
- Anthropic redacted most of the strongest evidence; safety lead Sam McCandlish said this was to avoid "handing out a roadmap for others to go faster."
- The report revealed two internal models ahead of Claude Opus 5.5, including a "Model 4" that cleared 50% on all 15 tested benchmarks, and METR found no clear bottlenecks.
OpenAI releases GPT-6 Cyber, a model for finding and exploiting software flaws
A frontier lab is now shipping a dedicated offensive-security model that outperforms its general flagship on cyber tasks.
- GPT-6 Cyber is built for hunting and exploiting software vulnerabilities and beats OpenAI's main GPT-6 Astra model on most cyber tasks.
- In testing, the model found 81 serious new bugs in Firefox's browser.
Benchmarks
| Benchmark | GPT-6 Astra | Claude Fable 5.2 | Gemini 4 Carbon | Muse Wonder | Kimi K3.1 Pro | Δ vs best other |
|---|---|---|---|---|---|---|
| GPQA diamond | 95.8 | 96.0 | 96.5 | — | 94.2 | -0.7 |
| FrontierMath-Tiers-1-3-v2-Private | 93.7 | 95.1 | 90.5 | 82.1 | 78.2 | -1.4 |
| FrontierMath-Tier-4-v2-Private | 97.6 | 100.0 | 95.1 | 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 | 75.6 | 94.0 | 77.7 | — | 62.2 | -18.4 |
| Chess Puzzles | 76.0 | 79.0 | 78.0 | — | 45.0 | -3.0 |
| EBR-bench | 76.2 | 90.5 | 87.6 | 31.4 | 37.1 | -14.3 |
| Mystery Game Puzzles | 84.0 | 90.0 | 73.0 | 51.0 | 42.0 | -6.0 |
| MirrorCode | 60.0 | — | — | 45.6 | 43.3 | +14.4 |
| ARC-AGI | 98.5 | 98.5 | 98.5 | — | — | +0.0 |
| Balrog | 68.3 | — | — | — | — | — |
| GSO-Bench | 79.4 | — | — | — | — | — |
| SimpleBench | 86.5 | 94.9 | 92.6 | — | — | -8.4 |
| WeirdML | 93.6 | — | — | — | — | — |
| APEX-Agents | 64.7 | 80.1 | 84.2 | 66.1 | 65.5 | -19.5 |
| ARC-AGI-2 | 95.0 | 93.8 | 94.2 | — | — | +0.8 |
| PostTrainBench | 41.9 | 44.6 | 41.0 | 24.2 | — | -2.7 |
| Remote Labor Index | 20.8 | 41.2 | — | — | — | -20.4 |
| ProofBench | 99.0 | 100.0 | — | 84.0 | 91.0 | -1.0 |
| FrontierCode | 53.3 | — | — | — | — | — |
| LMCA | 64.4 | 70.6 | 65.7 | 56.8 | 57.6 | -6.2 |
| DTBench | 97.3 | 99.2 | 98.4 | 95.2 | 93.3 | -1.9 |
| Furniture Assembly | 80.0 | 90.0 | 77.5 | 54.2 | 55.8 | -10.0 |
| FrontierSWE | 65.5 | 69.9 | 61.2 | — | — | -4.4 |
| WeirdML v3 | 54.0 | 65.3 | — | 24.9 | 18.9 | -11.3 |
| SpireClimb | 34.0 | 15.3 | — | — | 0.0 | +18.7 |
Δ = GPT-6 Astra minus the best other model listed. Sources: epoch.ai · Data: Epoch AI (CC BY 4.0)
Thinking Machines releases Inkling 2 Small, a 594B MoE with Apache 2.0 weights coming
A large mixture-of-experts model with vision and agentic building skills is heading to open weights after safety testing.
- Inkling 2 Small is a Mixture-of-Experts model with 594B total parameters and 17B active, supporting image inputs and variable effort.
- Its vision and agentic capabilities let it use visual feedback to iteratively build and refine websites, games and 3D designs with interactive elements.
- Initially available through Tinker with closed weights; full weights arrive under Apache 2.0 next week after external safety testing and red-teaming.
Benchmarks
| Benchmark | Inkling | GPT-6 Astra | Claude Opus 5.5 | Gemini 4 Argon | Muse Wonder | Δ vs best other |
|---|---|---|---|---|---|---|
| GPQA diamond | 88.3 | 95.8 | 90.6 | 96.4 | — | -8.1 |
| FrontierMath-Tiers-1-3-v2-Private | 33.3 | 93.7 | 91.2 | 86.0 | 82.1 | -60.4 |
| FrontierMath-Tier-4-v2-Private | 4.9 | 97.6 | 95.0 | 78.0 | 70.7 | -92.7 |
| OTIS Mock AIME 2024-2025 | 88.9 | 100.0 | 100.0 | 100.0 | 99.4 | -11.1 |
| SimpleQA Verified | 40.3 | 75.6 | 72.2 | 76.0 | — | -35.7 |
| Chess Puzzles | 21.0 | 76.0 | 67.0 | 59.0 | — | -55.0 |
| ARC-AGI | 79.5 | 98.5 | 98.5 | 98.5 | — | -19.0 |
| SimpleBench | 50.0 | — | — | — | — | — |
| WeirdML | 32.3 | 93.6 | — | — | — | -61.3 |
| APEX-Agents | 33.8 | 64.7 | 73.5 | 82.4 | 66.1 | -48.6 |
| ARC-AGI-2 | 36.5 | 95.0 | 93.3 | 91.2 | — | -58.5 |
| ProofBench | 0.0 | 99.0 | 100.0 | 99.0 | 84.0 | -100.0 |
| FrontierCode | 14.0 | 53.3 | 54.6 | 47.5 | — | -40.6 |
| LMCA | 37.6 | 64.4 | 68.2 | 63.7 | 56.8 | -30.6 |
| DTBench | 87.5 | 97.3 | 98.9 | 97.6 | 95.2 | -11.4 |
| FrontierSWE | 4.1 | 65.5 | 62.3 | 55.0 | — | -61.4 |
| WeirdML v3 | 4.4 | 54.0 | 49.8 | 41.1 | 25.2 | -49.6 |
Δ = Inkling minus the best other model listed. Sources: epoch.ai · Data: Epoch AI (CC BY 4.0)
OpenAI ships Images 2.6 with GIF generation
OpenAI's image model family adds motion for the first time while getting faster.
- Images 2.6 rolls out across ChatGPT and Codex, and can animate a typed idea, photo, or sketch into a GIF, or create animations from scratch.
- Two API versions are available to developers: Sunburst for higher-quality work and Flare for lighter, faster jobs.
- OpenAI says the model generates images 50% faster than previous versions, and 2.6 edges past its predecessor for the top spot on Design Arena.
Google rolls out Gemini Omni 1.2 Flash with a 30% price cut
Google's video model gets cheaper and generates longer clips.
- Pricing drops 30% to $0.07 per second at 720p.
- Maximum clip length extends from 10 seconds to 15 seconds.
Anthropic adds a Personal Finance tool to Claude
Claude can now connect directly to a user's financial accounts.
- The new tool lets Claude connect to a user's financial accounts, answer questions about them, and provide tailored information.
Ai2 opens ToyBlocks, a controlled simulation world for robot learning and evaluation
A deliberately narrow block-manipulation world lets researchers trace why robot policies fail instead of only scoring success rates.
- ToyBlocks auto-generates tasks, skills and demonstrations without teleoperation, running hundreds of plans in parallel on a single GPU at up to 24x real-time; the initial release includes 85,336 long-horizon demonstrations, about 1,600 hours of execution data and 144 million frames.
- Generalization tests found that ManiFlow, π0.5 and MolmoAct 2 generalize poorly to new spatial configurations despite large-scale pretraining; a hierarchical Molmo 2 + ManiFlow system scored 55.5% in simulation and 50.0% zero-shot on a real Franka FR3 arm.
- In agentic tests, a system using Claude Opus 5.5 over MCP tools solved 30/30 classical block-stacking tasks but reached only 63% on search-and-spell, with errors from look-alike letters (13%) and wrong letter orientation (15%); agent-written skill descriptions raised success from 24% to 63%, beating a human expert's 36%.
MIT shows GPT-6 Astra improving another model by editing its weights directly
A concrete, if narrow, demonstration of a model upgrading another model without any training run.
- Alibaba's small Qwen model scored 7.9% in math until Astra changed 0.016% of its weights to cut a repetition habit, reaching 48.1%.
- The same approach barely moved math scores for Google's smaller Gemma 3, which lacked a single flaw like Qwen's repetition.
- In smaller tests Astra named the game a model had learned from its weights alone and wrote a simple Breakout player from scratch; TLDR also flags commentary arguing the "read and write weights" framing is misleading and not superhuman.
OpenAI reports validating and patching 948 Linux privilege-escalation bugs in one run
The first large-scale result from OpenAI's autonomous vulnerability pipeline behind GPT-6 Cyber.
- OpenAI developed an autonomous pipeline for discovering, exploiting and patching security vulnerabilities.
- In a single run, its cybersecurity model validated and generated patches for 948 Linux local privilege escalation vulnerabilities.
| Benchmark | GPT-6 Astra | Inkling 2 Small | Muse Spark Open | Kimi K3.1 Pro | Claude Fable 5.2 | Gemini 4 Carbon | DeepSeek V4.1 Flash 1017 | Grok 4.6 |
|---|---|---|---|---|---|---|---|---|
| GPQA diamond | 95.8 | — | 88.6 | 94.2 | 96.0 | 96.5 | 93.2 | 94.0 |
| FrontierMath-Tiers-1-3-v2-Private | 93.7 | — | 69.8 | 78.2 | 95.1 | 90.5 | 71.2 | 66.0 |
| FrontierMath-Tier-4-v2-Private | 97.6 | — | 31.7 | 65.9 | 100.0 | 95.1 | 46.3 | 31.7 |
| OTIS Mock AIME 2024-2025 | 100.0 | — | 98.1 | 100.0 | 99.7 | 100.0 | 99.4 | 99.2 |
| SimpleQA Verified | 75.6 | — | 56.6 | 62.2 | 94.0 | 77.7 | 42.5 | 49.3 |
| Chess Puzzles | 76.0 | — | — | 45.0 | 79.0 | 78.0 | 46.0 | 33.0 |
| EBR-bench | 76.2 | — | — | 37.1 | 90.5 | 87.6 | — | 30.5 |
| Mystery Game Puzzles | 84.0 | — | — | 42.0 | 90.0 | 73.0 | 49.0 | 34.0 |
| MirrorCode | 60.0 | — | — | 43.3 | — | — | — | 41.1 |
| ARC-AGI | 98.5 | 96.5 | 89.0 | — | 98.5 | 98.5 | 94.5 | 87.5 |
| Balrog | 68.3 | — | — | — | — | — | — | — |
| GSO-Bench | 79.4 | — | — | — | — | — | — | — |
| SimpleBench | 86.5 | — | — | — | 94.9 | 92.6 | — | 75.9 |
| WeirdML | 93.6 | — | — | — | — | — | — | 67.3 |
| APEX-Agents | 64.7 | 46.8 | 33.4 | 65.5 | 80.1 | 84.2 | 69.8 | 65.3 |
| ARC-AGI-2 | 95.0 | 81.7 | 57.5 | — | 93.8 | 94.2 | 78.1 | 67.1 |
| PostTrainBench | 41.9 | — | — | — | 44.6 | 41.0 | — | — |
| Remote Labor Index | 20.8 | — | — | — | 41.2 | — | — | — |
| ProofBench | 99.0 | — | — | 91.0 | 100.0 | — | — | 51.0 |
| FrontierCode | 53.3 | — | — | — | — | — | — | 48.0 |
| LMCA | 64.4 | — | — | 57.6 | 70.6 | 65.7 | — | 48.5 |
| DTBench | 97.3 | — | — | 93.3 | 99.2 | 98.4 | — | 97.3 |
| Furniture Assembly | 80.0 | — | — | 55.8 | 90.0 | 77.5 | — | 40.0 |
| FrontierSWE | 65.5 | 25.1 | — | — | 69.9 | 61.2 | — | 25.3 |
| WeirdML v3 | 54.0 | — | — | 18.9 | 65.3 | — | 19.7 | — |
| SpireClimb | 34.0 | — | — | 0.0 | 15.3 | — | — | — |
Highlighted: released in this period. Bold: best score. Independent scores from Epoch AI (CC BY 4.0).