The short version
- Claude Fable 5.2 is generally available and about 30% faster than the previous version. Mythos 5.2, the same model, is limited to vetted cyber-defence and life-science organisations.
- Anthropic says 1,000 Fable 5.2 agents fully formalised Fermat's Last Theorem in 30 hours. That is a machine-checked translation of the existing proof, not a new proof.
- Meta's Muse Ember costs $0.10 per million input tokens with a 1M-token context, aimed at cheap, repeatable agent work.
- Reports say Anthropic took on $3.4 billion of gas generators, and China unveiled the Xingyun space-computing constellation. Power is becoming the limit on AI capacity.
November 10 covered both ends of the model market. Anthropic released its most capable generally available model, and Meta released its smallest and cheapest. Behind both, AI labs are searching for electricity and computing capacity in unusual places, from gas generators in Texas to satellites in orbit. This issue explains each release and how to decide which class of model a task needs.
01
Claude Fable 5.2 and Mythos 5.2: a faster frontier model with restricted access
What Anthropic released
Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9. The company says both generate output about 30% faster than their predecessors, with gains in coding, reasoning and vision. Fable 5.2 is generally available on paid Claude plans, in Claude Code, through the Claude API and on AWS, Google Cloud and Microsoft Azure. Mythos 5.2 is the same model, available only to organisations verified through Anthropic's trusted-access programmes for cyber defence and life sciences. Its listed API price starts at $8 per million input tokens and $40 per million output tokens.
Anthropic published two research demonstrations with the launch. In the first, a team of 1,000 Fable 5.2 agents fully formalised Fermat's Last Theorem in 30 hours. They translated the existing proof into a form a computer can check step by step. Under the same setup, Opus 5.5 completed 26% in 18 hours. The agents did not discover a new proof. In the second, Mythos 5.2 combined existing protein-structure tools to model viral capsids, the protein shells of viruses, and Anthropic reports accurate models for 28 of the 34 it attempted. Both are company-reported results.
When a frontier model is worth it
The formalisation result shows what the top tier is for: long, multi-step work where the output can be checked. Most daily tasks do not need it. Emails, summaries and routine code changes are usually handled well and more cheaply by a smaller model.
Use a simple rule. If a task takes you more than an hour, has a clear way to verify the result, and a mistake would be costly, try the frontier model. Examples are a large refactor with a test suite, a data migration with validation queries, or a long contract comparison against a checklist. Otherwise, start with your provider's faster tier. In practice, editors such as Cursor and GitHub Copilot let you choose the model per task, so you can keep Claude Fable 5.2 for the hard tickets only. Track two numbers for a week: time to a passing result and cost per task. Our Cursor vs GitHub Copilot comparison covers the editors themselves.
Get HAI Daily in your inbox
The day's AI news, what it changes, and which tools to try.
02
Meta Muse Ember: a very cheap model for agent work
What Meta launched
Meta introduced Muse Ember on November 9 as its smallest and most efficient model. It costs $0.10 per million input tokens and $0.50 per million output tokens, with a 1-million-token context window and up to 512,000 output tokens. Prompts over 200,000 tokens are charged at a higher rate. Ember understands text, images, audio and video, so an agent can read screenshots, charts and scanned documents directly. It is available through the Meta Model API and Muse Code.
Meta says Ember matches the best score of Anthropic's Claude Haiku 5.5 on its comparison at about a third of the cost per task. Those per-task figures are Meta's estimates. One analysis noted that Meta's chart and launch text use slightly different values.
Where a budget model fits
Cheap models earn their place in the repetitive steps of a workflow: classifying incoming messages, extracting fields from documents, checking that a form is complete, or summarising a screenshot before a stronger model decides what to do. In an agent, these steps often run hundreds of times, so price per call matters more than peak capability.
Test a budget model on a sample of real items before switching. Take 50 recent documents or messages, run them through the cheap model and through your current model, and count disagreements. If the cheap model agrees on almost all of them and the disagreements are easy to catch, route that step to it and keep the stronger model for exceptions. Meta's models are not yet listed in our directory. To compare general assistants, start in the Chatbots category with Gemini or ChatGPT, both of which also offer low-cost tiers.
03
Power and space: where AI labs are finding capacity
Generators in Texas, satellites in orbit
According to a Financial Times report on November 8, Anthropic took over commitments for 1.7 gigawatts of heavy-duty gas generators worth about $3.4 billion from OnePWR Solutions, an equipment supplier in San Antonio, Texas. The deal reportedly includes the engineering and construction capability to install them. The FT called it the largest known purchase of power equipment by an AI lab. Neither company commented, and Reuters said it could not immediately verify the report, so the figures should be treated as reported rather than confirmed.
In China, the Aerospace Information Research Institute of the Chinese Academy of Sciences and domestic technology companies presented the Xingyun Constellation Plan at a space-cloud industry conference on November 9. It is a plan for a computing network in orbit built on Chinese chips, servers and satellites. It is meant to provide computing for natural-resource monitoring, disaster response and low-altitude aviation. Satellite numbers and launch dates were not published.
Why this matters to people who buy AI services
Both stories show that electricity and physical capacity, not only chips, now limit how much AI service providers can offer. When capacity is tight, providers ration access to their most demanding models, raise prices or set lower usage limits. When new capacity arrives, those limits ease.
For buyers, the practical response is to plan for limits. Know the rate limits and usage caps on your current plan, and keep a second provider tested for important work. If you follow AI infrastructure companies for research, company databases such as Crunchbase AI in the Finance AI category help track deals. Check each figure against the original report, especially when a deal is unconfirmed, as the generator purchase is.
What to try next
Match the model to the task
Use frontier models such as Fable 5.2 for long tasks you can verify. Use budget models such as Muse Ember for high-volume steps you can sample and check. Keep a mid-tier model for everything in between. Keep a second provider ready, because capacity and pricing will keep changing as labs compete for power.
Compare options in Chatbots and Code Generation, and read our ChatGPT vs Claude guide for a broader comparison.
How we put this issue together
The H.AI Tools Team checks each story against the official sources linked in the text and removes claims we could not confirm. Product links lead to H.AI Tools directory profiles. Unless an issue says otherwise, product notes describe vendor documentation, not our own benchmark tests.