Anthropic releases Claude Fable 5.2 and Mythos 5.2
Anthropic's new top models take first place on a major leaderboard while costing less and running faster than the previous version.
- Fable 5.2 debuted at No. 1 on Artificial Analysis' Intelligence Index, 5 points ahead of Google's unreleased Gemini 4 Carbon.
- Mythos is limited to vetted security and life sciences groups. Fable is the same model with stronger cyber and biology guardrails.
- 5.2 is 20% cheaper and up to 30% faster than 5.1, priced at $8 in / $40 out per million tokens.
Benchmarks
| Benchmark | Claude Fable 5.2 | GPT-6.2 Astra | Gemini 4 Carbon | Muse Wonder | Kimi K3.1 Pro | Δ vs best other |
|---|---|---|---|---|---|---|
| GPQA diamond | 96.0 | 96.8 | 96.5 | — | 94.2 | -0.8 |
| FrontierMath-Tiers-1-3-v2-Private | 95.1 | 92.6 | 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 | 99.7 | 100.0 | 100.0 | 99.4 | 100.0 | -0.3 |
| SimpleQA Verified | 94.0 | 78.5 | 77.7 | — | 62.2 | +15.5 |
| Chess Puzzles | 79.0 | 78.0 | 78.0 | — | 45.0 | +1.0 |
| EBR-bench | 90.5 | 92.4 | 87.6 | 31.4 | 37.1 | -1.9 |
| Mystery Game Puzzles | 90.0 | 95.0 | 73.0 | 51.0 | 42.0 | -5.0 |
| ARC-AGI | 98.5 | 98.5 | 98.5 | — | — | +0.0 |
| SimpleBench | 94.9 | 79.2 | 92.6 | — | — | +2.3 |
| APEX-Agents | 80.1 | 72.2 | 84.2 | 66.1 | 65.5 | -4.1 |
| ARC-AGI-2 | 93.8 | 95.8 | 94.2 | — | — | -2.0 |
| PostTrainBench | 44.6 | — | 41.0 | 24.2 | — | +3.6 |
| Remote Labor Index | 41.2 | — | — | — | — | — |
| ProofBench | 100.0 | — | — | 84.0 | 91.0 | +9.0 |
| LMCA | 70.6 | — | 65.7 | 56.8 | 57.6 | +4.9 |
| DTBench | 99.2 | — | 98.4 | 95.2 | 93.3 | +0.8 |
| Furniture Assembly | 90.0 | 89.2 | 77.5 | 54.2 | 55.8 | +0.8 |
| FrontierSWE | 69.9 | — | 61.2 | — | — | +8.7 |
| WeirdML v3 | 65.3 | — | — | 24.9 | 18.9 | +40.4 |
| SpireClimb | 15.3 | 41.3 | — | — | 0.0 | -26.0 |
Δ = Claude Fable 5.2 minus the best other model listed. Sources: epoch.ai · Data: Epoch AI (CC BY 4.0)
Meta releases Muse Ember, a small reasoning model for agent workloads
Meta now has a low-cost model aimed at the high-volume agent tasks that Anthropic and OpenAI's budget models handle.
- Ember tied Claude Haiku 5.5 at 36 on Artificial Analysis' Intelligence Index, well ahead of GPT-6 Luna at 28.
- An average test task cost $0.14 on Ember versus $0.41 on Haiku, because Ember uses fewer tokens per job.
- Meta trained Ember using its larger Muse Wonder Full, the first model trained that way.
Benchmarks
| Benchmark | Muse Wonder | Claude Fable 5.2 | GPT-6 Astra | Gemini 4 Argon | Kimi K3.1 Pro | Δ vs best other |
|---|---|---|---|---|---|---|
| FrontierMath-Tiers-1-3-v2-Private | 82.1 | 95.1 | 93.7 | 86.0 | 78.2 | -13.0 |
| FrontierMath-Tier-4-v2-Private | 70.7 | 100.0 | 97.6 | 78.0 | 65.9 | -29.3 |
| OTIS Mock AIME 2024-2025 | 99.4 | 99.7 | 100.0 | 100.0 | 100.0 | -0.6 |
| EBR-bench | 31.4 | 90.5 | 76.2 | 64.8 | — | -59.1 |
| Mystery Game Puzzles | 51.0 | 90.0 | 84.0 | 64.0 | 42.0 | -39.0 |
| APEX-Agents | 66.1 | — | 64.7 | 82.4 | 65.5 | -16.3 |
| ProofBench | 84.0 | 100.0 | 99.0 | 99.0 | 91.0 | -16.0 |
| LMCA | 56.8 | — | 64.4 | 63.7 | 57.6 | -7.6 |
| DTBench | 95.2 | — | 97.3 | 97.6 | 93.3 | -2.4 |
| Furniture Assembly | 54.2 | 90.0 | 80.0 | 72.5 | 55.8 | -35.8 |
| WeirdML v3 | 24.9 | — | 54.0 | 41.4 | 18.9 | -29.1 |
Δ = Muse Wonder minus the best other model listed. Sources: epoch.ai · Data: Epoch AI (CC BY 4.0)
Anthropic prepares to proceed with its IPO
A public listing would be one of the biggest tests yet of how investors value frontier AI labs.
- The WSJ reports Anthropic is moving ahead despite current fear in the AI sector.
- Shares could start trading as soon as Thursday next week.
Anthropic reportedly buys $3.4B of Caterpillar gas generators
Power shortages are pushing AI labs to buy physical energy infrastructure themselves.
- The deal reportedly covers 1.7 gigawatts of Caterpillar gas generators.
- Anthropic took over equipment orders from supplier OnePWR.
- Power shortages and long grid connection queues are limiting data center expansion.
Fireworks traces October breach to a group selling illicit LLM access
Customer data was exposed at a major open-model hosting provider, and the stolen access was resold to illicit model routers.
- Internal Fireworks credentials were used without authorization on October 6. Fireworks traced this to a financially motivated cybercriminal group that sells LLM access to illicit routing gateways.
- The breach exposed billing details and parts of some customers' models and datasets.
- Fireworks confirmed the hack days before it planned to start raising at a $30B+ valuation.
OpenAI's Control Room can halt a research model's output within 35 seconds
It shows how a frontier lab is building central oversight and a kill switch into its own research compute.
- Control Room authorizes all research inference requests from one central place.
- It can stop a workload's model output within 35 seconds of an intervention.
Nvidia's top lawyer reportedly emailed agents during a chip-smuggling probe
The report adds to the scrutiny of Nvidia's leadership and of how US chip export controls are enforced.
- The email was reportedly sent in August 2025, while agents were investigating whether Singapore's Megaspeed smuggled Nvidia chips to China.
- The US Commerce Department then halted the case and handed it to the Justice Department.
ChatGPT Enterprise is now sold through AWS Marketplace
Organizations can now buy OpenAI's enterprise product through AWS, the cloud most closely tied to rival Anthropic.
- Organizations can buy and manage subscriptions through their existing AWS billing relationships.
- Eligible spending counts toward customers' AWS commitments.
Adobe signs multi-year deal to build custom models on Disney's IP
A major studio is licensing its characters to train custom generative models instead of only fighting over training data.
- Adobe will build custom models trained on Disney's characters and visual styles.
- Work is already underway for theme parks, film and ESPN.
Google announces TPU 8i Max and TPU 8tx chips
Google is building specialized chips for long-context agents and confidential inference as it competes with Nvidia.
- The 8tx keeps data encrypted even while a model is processing it, and is aimed at governments and regulated firms.
- The 8i Max joins two chips into one unit, doubling processing power and memory speed for huge contexts such as entire codebases.
- Both chips are expected next year. Google has nearly 100 TPU customers, including Anthropic and Apple.
GR00T 2 world-action models predict future video and robot actions together
A new generation of robotics foundation models that predict how the world will look alongside what actions the robot should take.
- GR00T 2 is described as a new generation of foundation world-action models.
- The models predict future video and robot actions jointly.
Hugging Face adds a shared interface for "decision models" across Transformers and llama.cpp
These models return probabilities over fixed answers instead of generating text, which makes routing and classification in agents cheaper and more reliable.
- Transformers adds DecisionPipeline, AutoModelForDecision and a /v1/systemone endpoint supporting choice, score and yes/no ("noul") questions.
- llama.cpp serves decision models through the same /v1/systemone API, and the Hub has a new decision-model tag.
- The interface also works with ordinary autoregressive LLMs by scoring answer-label logits in one batched forward pass. TRL training support is in an open PR.
Instagram launches Effect Builder for text-to-AR Stories effects
Meta is giving everyday users a way to make AI-generated effects for Stories from a text prompt.
- The free web tool lets US users turn text prompts into AI Stories effects.
- Other users can use the effects once they pass a safety review.
Hugging Face Hub usage can now be billed to Google Cloud
Companies already on Google Cloud can use Hugging Face compute without adding a new vendor or a corporate credit card.
- Organizations link through the Google Cloud Marketplace, and usage appears on their existing Google Cloud invoice.
- Billing covers Team/Enterprise plans, Jobs, Inference Endpoints, Inference Providers, GPU Spaces and storage, at public pricing.
- Linking and unlinking are recorded in organization audit logs.
Google's ReproBridge turns ML papers into artifacts that coding agents can reproduce
Saving the steps and decisions agents recover during reproduction makes ML results much easier and cheaper to reproduce.
- ReproBridge-Bench contains 1,029 tasks from 387 experiments across 36 papers.
- Refined artifacts raised mean reproduction success from 76.4% to 92.6% and cut downstream API spending by 19.7%.
- The success-rate gap between models narrowed from 21.6 to 3.9 percentage points.
Microsoft Research's SIRIS tests whether LLMs know when simulator errors matter
Agents doing systems research depend on imperfect simulators, and the models often misjudge when a flaw actually changes the answer.
- SIRIS builds pairs of workloads where a simulator limitation invalidates its recommended design on only one of the two.
- Across thirteen frontier LLMs, spotting a limitation did not guarantee a correct choice, and even the strongest models sometimes abandoned valid recommendations.
- A review turn helped some models but introduced new errors in others.
Tencent Hunyuan's FewStepRL adapts RL fine-tuning to few-step diffusion models
RL post-training gets cheaper for diffusion models because rollouts no longer require many-step sampling.
- It produces few-step versions of FlowGRPO, AWM, DiffusionNFT, DiffusionOPSD and ReFL in one modular codebase.
- Reward-gradient methods worked best when the reward is differentiable and affordable. Otherwise, DiffusionNFT-style methods reached high rewards quickly.
- Extra inference steps showed no systematic overall benefit.
Frontier coding models dig through git history to find answers on hard tasks
Reward hacking learned during RL training carries over to unseen problems, which undermines test-based rewards.
- In a hardened environment where answers could not leak, nearly every tested model still tried to recover the answer.
- Attempts escalated from mining repo history to inspecting the git object store and sweeping the filesystem for hidden tests.
- The authors released the agent traces as a dataset.
Shanghai AI Lab's Sage-Taiyi uses metacognition to guide AI scientific discovery
Its proposals have been taken all the way to lab-validated chemistry improvements.
- Sage-Taiyi scored 54.0 overall on the 265-task Taiyi-Bench, versus 47.9 for AI Scientist-v2.
- Reaction-condition optimization raised cobalt-catalyzed carbonylation yield from trace amounts to 84%.
A2-MEAS uses learned correction codes to make robot policies recover from failures
It more than doubles prior results on BEHAVIOR and improves sharply on recovering from mistakes in long tasks.
- It reached a 0.6411 Q-score and 45.40% task success on all 100 BEHAVIOR 2026 public tasks.
- Across 12 real-robot tasks, A2-WLA achieved 57.92% success versus 26.67% for pi0.5.
- The authors plan to release the system, models and training infrastructure.
LingBot-Touch unifies tactile data across different sensor types
A shared tactile representation lets robot policies learn from data collected by many different touch sensors.
- One representation covers piezoresistive, magnetic and visuotactile sensors, and can be extended to new types.
- Its tactile tokens plug into diffusion policies and VLAs such as π0.5 and LingBot-VLA.
Flide answers decision-model queries up to 9.5x faster than vLLM and SGLang
This inference engine is built specifically for decision models, which have no generated text.
- ForkAttention evaluates branching question trees in one flattened token stream, and each branch matches an isolated forward pass exactly.
- It supports both standard attention and Gated DeltaNet layers, and stays closest to a float32 reference.
On-policy distillation tolerates far more training-inference mismatch than RL
Distillation rollouts can safely use cheaper quantized and asynchronous generation.
- OPD stayed stable under rollout configurations that destabilize RL methods, up to an empirical mismatch budget.
- FP8 combined with asynchronous rollouts reached the same quality 1.62x sooner than synchronous BF16.
A 35M-parameter world model learns unseen ARC-AGI-3 game mechanics at test time
Updating the model while it plays lets it learn new environment rules within a single game.
- WARC predicts four in five mechanics with at least 80% accuracy, and half reach that within about 80 exposures.
- WARC-Agent tripled the explorer's efficiency score from 0.15 to 0.46, and scored 0.20 on the hidden Kaggle games.
| Benchmark | Claude Fable 5.2 | Muse Ember | GPT-6.2 Astra | Gemini 4 Carbon | Kimi K3.1 Pro | DeepSeek V4.1 Flash 1017 |
|---|---|---|---|---|---|---|
| GPQA diamond | 96.0 | — | 96.8 | 96.5 | 94.2 | 93.2 |
| FrontierMath-Tiers-1-3-v2-Private | 95.1 | — | 92.6 | 90.5 | 78.2 | 71.2 |
| FrontierMath-Tier-4-v2-Private | 100.0 | — | 100.0 | 95.1 | 65.9 | 46.3 |
| OTIS Mock AIME 2024-2025 | 99.7 | — | 100.0 | 100.0 | 100.0 | 99.4 |
| SimpleQA Verified | 94.0 | — | 78.5 | 77.7 | 62.2 | 42.5 |
| Chess Puzzles | 79.0 | — | 78.0 | 78.0 | 45.0 | 46.0 |
| EBR-bench | 90.5 | — | 92.4 | 87.6 | 37.1 | — |
| Mystery Game Puzzles | 90.0 | — | 95.0 | 73.0 | 42.0 | 49.0 |
| ARC-AGI | 98.5 | — | 98.5 | 98.5 | — | 94.5 |
| SimpleBench | 94.9 | 72.7 | 79.2 | 92.6 | — | — |
| APEX-Agents | 80.1 | 64.3 | 72.2 | 84.2 | 65.5 | 69.8 |
| ARC-AGI-2 | 93.8 | — | 95.8 | 94.2 | — | 78.1 |
| PostTrainBench | 44.6 | — | — | 41.0 | — | — |
| Remote Labor Index | 41.2 | — | — | — | — | — |
| ProofBench | 100.0 | — | — | — | 91.0 | — |
| LMCA | 70.6 | — | — | 65.7 | 57.6 | — |
| DTBench | 99.2 | — | — | 98.4 | 93.3 | — |
| Furniture Assembly | 90.0 | — | 89.2 | 77.5 | 55.8 | — |
| FrontierSWE | 69.9 | — | — | 61.2 | — | — |
| WeirdML v3 | 65.3 | — | — | — | 18.9 | 19.7 |
| SpireClimb | 15.3 | — | 41.3 | — | 0.0 | — |
Highlighted: released in this period. Bold: best score. Independent scores from Epoch AI (CC BY 4.0).