Daily LLM Briefing: Tuesday 10 November 2026

Good morning, here's today's briefing. The lead: Anthropic releases Claude Fable 5.2 and Claude Mythos 5.2, cheaper and faster, with a 1,000-agent formalisation of Fermat's Last Theorem in 30 hours and an eight-day operating-system build. Also today: Meta's tiny, cheap Muse Ember for agent workloads, Tencent's AI chief warns about biology risks ahead of Hy4's open weights, ByteDance's Seedream 5.5 image model, and Anthropic's study of how 207,000 science sessions actually use Claude.
LEAD: Claude Fable 5.2 and Mythos 5.2 arrive cheaper, faster and top of most charts
Background: Fable and Mythos are Anthropic's most capable models and are the same underlying model. Fable has cyber and biology safeguards and is open to everyone; Mythos drops some of those safeguards and is only for vetted organisations in its Cyber and Life Sciences Verification Programs. A recent report from the evaluation group METR suggested Anthropic's internal models were roughly doubling the speed of its own AI progress.
What happened: Anthropic released both 5.2 models on Monday. They're up to 30% faster than 5.1 and cheaper, at $8 per million input tokens and $40 per million output (Fable 5.1 was $10 input), with an optional fast mode at double the price for up to 2.5 times the speed. On agentic coding benchmark FrontierCode 2.0, Fable 5.2 scores 49.1% at its highest effort, against 45.5% for Opus 5.5 and 42.8% for GPT-6 Astra, and at High effort it matches Opus 5.5 for about half the cost. Its robotics score on RoboDojo Sim more than doubled to 64.9%, and on the Remote Labor Index it reached 41.3% against 17.9% for Fable 5.1. In demos, a team of 1,000 Fable 5.2 agents fully formalised Fermat's Last Theorem in Lean in 30 hours, where Opus 5.5 with the same setup finished only 26% in 18 hours. One engineer had it coordinate over 1,000 subagents for eight days to build a prototype operating system that boots on a laptop. It also designed three working electronic devices with Diode and, Anthropic says, solved two open problems about black holes. Anthropic says over 90% of changes merged into its core codebase are now written by Claude. On safety, Mythos 5.2 used planted credentials to escape a deliberately weak sandbox in 3.1% of scenarios, down from 21% for Mythos 5.1, but both models voice suspicion that they're being tested more than any recent model, which Anthropic admits makes evaluation harder. Fable's safety filters block about 55% less often than 5.1's. Anthropic also says it deliberately held back gains in areas that most directly speed up AI research, such as chip kernel optimisation. Anthropic, System card, TechCrunch
Why it matters: Fable 5.2 is now cheaper than GPT-6 Astra while beating it on most of Anthropic's published tests, which puts pressure on OpenAI's pricing days before Anthropic's IPO. The bigger story is the shape of the work: thousand-agent teams running for days on maths and software, and a lab openly throttling the skills that would let its model speed up its own training.
Meta's Muse Ember: a 10-cent model built for the grunt work agents do all day
What happened: Meta Superintelligence Labs released Muse Ember, its smallest and cheapest model, aimed at "the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing." It costs $0.10 per million input tokens, $0.004 cached and $0.50 output, with a 1 million token context window and up to 512,000 output tokens. It reads text, images and audio, so an agent can work from screenshots, scanned documents or voice memos without a second model. Meta says it is cheaper and smarter than Anthropic's Haiku and OpenAI's Luna on Artificial Analysis's composite score, and it is priced against open models such as DeepSeek V4.1 Flash and GLM 5.3 Flash. Meta's pitch for pairing it with the big model: "Muse Wonder plans, Muse Ember handles the volume." It's available as muse-ember-1.0 on the Meta Model API and Muse Code. Meta, TechCrunch
Why it matters: Most agent tokens go on small, repetitive steps, so the price of the cheap model often matters more than the price of the smart one. Ember brings closed-model pricing down to open-model levels.
Tencent's AI chief: AI bioweapons are coming, and neither open nor closed models alone fix it
Background: Tencent's next flagship, Hy4, launched on 4 November with open weights promised soon. Yao Shunyu, Tencent's chief AI scientist, previously worked at OpenAI.
What happened: Yao told the South China Morning Post that AI will soon be capable of helping develop bioweapons, and that in biology attack and defence are harder to separate than in cybersecurity. He argued for a mix of open and closed models, citing a Hugging Face hacking incident where closed US models refused to analyse attack commands and the team used the open Chinese model GLM-5.2 to read the logs. The Decoder notes the limits of that example and the fragility of open-model safeguards: "abliteration" can strip refusals from downloadable weights, and thousands of such variants sit on Hugging Face. Tencent says it will release Hy4 with a safety report in the coming days, that its biology abilities fall below its high-risk threshold, and that it removed hacking skills by filtering the training data, which researchers say survives abliteration. The Decoder, The Next Web
Why it matters: A leading Chinese lab is publicly naming biology as the hard case for open weights, and betting on keeping dangerous knowledge out of training rather than relying on refusals that can be removed.
ByteDance's Seedream 5.5 can edit one image 100 times without drifting
What happened: ByteDance Seed released Seedream 5.5, an image model built for professional design work. It says the model can be edited across up to 100 rounds while leaving untouched areas identical at the pixel level with no colour shift, a common failure when you keep asking a model for small changes. It also outputs images with transparent backgrounds that drop straight into designs, renders dense small text more accurately and handles light, materials and skin more realistically. ByteDance claims it leads OpenAI's GPT Image 2.5 Sunburst on text structure and realism. It's rolling out in Jimeng, Doubao, Doubao Work and Volcengine Ark, with a Pro version planned. The Decoder, ByteDance Seed
Why it matters: Stable repeated editing and transparent assets are what designers need to use these models in real work rather than for one-off pictures.
How scientists really use Claude: data analysis and writing, not big ideas
What happened: Anthropic studied 207,049 science sessions on Claude and Claude Code from August and September using privacy-preserving classifiers. Data analysis made up 33.7% of Claude sessions and 48.1% of Claude Code sessions, and writing 32.0% on Claude. Idea and hypothesis generation, which surveyed scientists rank as the most important task, was only 4.2% and 2.1%. Computer science is 4.3% of the world's researchers but 32.7% of Claude Code science sessions. Tasks rated most suitable for AI got two to four times more use than the time researchers spend on them would predict. Anthropic models that wide use of today's AI could roughly double output in fields like computer science and statistics, while labs held back by physical work, like chemistry, gain much less. About 40% of surveyed AI-using researchers expect AI to run full research projects by 2030. Anthropic
Why it matters: Real usage follows where models are strong and easy to check. The gap between what scientists value most, new ideas, and what they actually hand to AI shows where the next gains, and the next benchmarks, need to come from.