The AI log, frontier · robotics · science
9 Nov 2026
Fable 5.2, Muse Ember, Skild's Steel solver, Language Explorer
Anthropic's Fable and Mythos 5.2, Meta's cheap agent model Muse Ember, a GPU physics solver for robot learning and Google's language atlas.
By Ali Asif, published
What shipped
- Anthropic
Introducing Claude Fable 5.2 and Claude Mythos 5.2
An upgrade to the same underlying model, with Fable on all platforms at $8 and $40 per million tokens and Mythos limited to verification programs, reporting 49.1% on FrontierCode 2.0 and 69.9% on FrontierSWE v2.
- Meta
Our Smallest, Most Efficient Model: Muse Ember
A small agent model with a 1M-token context window priced at $0.10 input and $0.50 output per million tokens, scoring 36 on the Artificial Analysis Intelligence Index.
- Skild AI
Steel: Scaling Simulation Across Robots, Tasks, and Worlds
An open source GPU rigid-body solver for the Newton physics engine that reaches 3.8x PhysX geometric-mean throughput across 20 tasks on one RTX PRO 6000 Blackwell.
- Google Research
Exploring the world's languages with Language Explorer
A free interactive platform built on the open source LinguaMeta dataset, part of the 1,000 Languages Initiative, also installed in a Berlin museum exhibition.
Four primary pages I opened and read today: a frontier model upgrade, a small and cheap agent model, a robotics simulation solver and a language atlas. Everything below comes from those pages, not from my own testing.
Claude Fable 5.2 and Mythos 5.2
Anthropic describes Fable 5.2 and Mythos 5.2 as the same underlying model, packaged two ways. Fable carries cybersecurity and biology safeguards and is available to all users on every major cloud, with the API ID claude-fable-5-2. Mythos goes to vetted organizations through the Cyber Verification Program and the Life Sciences Verification Program. Standard pricing is $8 per million input tokens and $40 per million output tokens, with a fast mode at $16 and $80. Anthropic says outputs are up to 30% faster than the 5.1 generation.
The reported benchmarks are all Anthropic's. FrontierCode 2.0 is 49.1% at Xhigh effort against 45.5% for Opus 5.5. FrontierSWE v2 is 69.9% against 62.3%. Humanity's Last Exam with tools is 74.2% against 67.7%, and the Remote Labor Index is 41.3% against 21.3%. Mythos 5.2 reaches 89.6% on CVE Rediscovery.
The safety section is the part worth reading slowly. On the automated behavioral audit Mythos 5.2 scored 2.67, where lower is better, against 3.10 for Opus 5.5. In the sandbox-escape evaluation it used outside credentials or files in 3.1% of scenarios, compared with 21% for Mythos 5.1. Anthropic also says the models suspect they are being evaluated more often than recent models, which makes the evaluations harder to read. The page lists the UK AI Security Institute and METR among external testers, and notes that the Fermat's Last Theorem formalization claim, done by 1,000 agents in about 30 hours, was not reviewed externally.
Meta's Muse Ember
Meta calls Muse Ember its smallest and most efficient model, built for tool calls, coding subtasks, routing, extraction and long document parsing. It has a 1M-token context window, up to 512K output tokens, and is on Muse Code and the Meta Model API as muse-ember-1.0, plus Google Cloud and Oracle Cloud. Pricing is $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 200K.
At max reasoning it scores 36 on the Artificial Analysis Intelligence Index, one point under Muse Spark 1.3, at $0.14 per task against $2.39. It beats Claude Haiku 5.5 on Terminal-Bench 4.0 (40.9% against 39.2%) and AutomationBench-AA (37.8% against 13.9%), but trails it on FrontierCode 2.0-AA, 27.2 against 34.9. The page gives no parameter count and states no license.
Steel from Skild AI
Steel is a GPU rigid-body dynamics solver from Skild AI and NVIDIA, contributed to the open source Newton Physics Engine as newton.solvers.SolverSteel and aimed at robot learning in Isaac Lab. On one RTX PRO 6000 Blackwell GPU, geometric-mean throughput across 20 single tasks is 3.8x PhysX and about 1.3x MuJoCo Warp. On mixed batches of 5 to 20 task types it reaches 9.2x to 13.0x PhysX's peak. At a matched batch of 61,440 environments it uses about 62% less GPU memory.
The caveat is stated on the page: these are simulation throughput figures, not end-to-end policy training speed.
Language Explorer
Google Research released Language Explorer, a free interactive platform for exploring linguistic diversity, built on the open source LinguaMeta dataset as part of the 1,000 Languages Initiative. It is also an installation in an exhibition at the Museum of Communication in Berlin. Useful for researchers and teachers, and small next to the rest of today.
What I'd watch next
Independent evaluations of Fable 5.2, since the claims so far are the lab's own, and whether Meta states a license and parameter count for Muse Ember. For Steel, I want to see a trained policy, not only a throughput chart.
Sources
Topics: Anthropic, NVIDIA, Google Research, Meta, Mythos, Fable