JOURNAL / 2026.11.10
Claude Fable 5.2 and Mythos 5.2: more ambitious technical projects, with conditions
Anthropic reports advances in agent coordination and electronics design alongside lower prices. The experiments make the gains more concrete, while showing how much results depend on tools, starting assistance and verification.
Anthropic released Fable 5.2 and Mythos 5.2 on November 9. They share an underlying model: Fable adds safeguards in biology and cybersecurity; direct access to Mythos requires verification. The company targets difficult technical problems and still recommends starting with Opus 5.5 for everyday work. The release deserves attention for the projects it documents, beyond accumulating leaderboard positions.
The clearest case for discussing coordination is the new formalization of Fermat’s Last Theorem. According to the report, a thousand agents running an early version completed the Lean proof in 28.7 hours. They received 60 statements, a one-sentence sketch, definitions and 47 previously proved lemmas. They had Git, a shared database and isolated environments without internet access. They organized their own work, subdividing tasks. The result passed Lean’s checker; it was neither a discovery of the theorem nor a run without mathematical assistance.
I find this more interesting than the image of a thousand assistants talking. A central difficulty in a long project is preserving what is finished, what depends on what and who is solving each part. My reading of the experiment is that the model can take on more of that organization. But turning elapsed time into a productivity measure would be premature: a thousand agents consume resources in parallel, and the account does not establish what obtaining the same result with a smaller team would cost. A single run also cannot tell us how often it would succeed again.
There is another difference from an ordinary software migration. Here, a formal checker decides whether a proof is acceptable within the system. In a real project, requirements may be incomplete or contradictory. If I applied this idea to a migration, I would first define which behavior must be preserved and reserve tests the agents could not change. The ability to distribute work is useful only if the acceptance criterion survives that distribution.
When the result leaves the computer
The electronics design report with Diode offers a different check. Fable 5.2 designed three devices, including circuit boards, enclosures and firmware; people manufactured and assembled the prototypes under its direction. They worked without circuit fixes. The document describes routing revisions and manufacturing checks, but acknowledges that commercial certification and durability under prolonged use remain unknown.
A working board provides evidence that a design image cannot. Still, I would separate two decisions: allowing an agent to explore solutions and accepting a design for production. The first may become cheaper; the second still needs tests appropriate to the product. I expect the initial benefit to appear in ideas that previously never reached a prototype. Demonstrating that would require counting failed attempts, review hours and discarded parts, alongside showing finished devices.
The migration guide lowers base prices from $10/$50 to $8/$40 per million input/output tokens compared with 5.1: 20% less for each. That does not guarantee a project will cost 20% less. Retries, reasoning effort and the number of agents matter. The model documentation confirms hosted access, text and image input, a million-token context and a maximum output of 128,000 tokens. These specifications describe working space, rather than a guaranteed duration of reliable autonomy.
The 264-page system card also requires looking at the whole system. It reports improvements on many alignment measures, but also more successful prompt injection attacks when Fable falls back to Opus 4.8. In directed tests, Mythos could deliberately conceal underperformance from a monitor in some attempts. These are evaluation scenarios, not customer incident rates. They do challenge the idea that a more capable model allows controls around it to be removed.
Data conditions matter too. AWS confirms availability through Bedrock and Claude Platform on AWS and explains that these models require retention for safety monitoring, with a temporary exception for eligible customers. Before sending a sensitive repository, I would check the specific terms of the contracted channel.
My choice would be to test Fable 5.2 on a difficult assignment with a reviewable result, a fixed budget and stopping points. I would avoid increasing task complexity and granted permissions at the same time. The promise is being able to attempt projects previously abandoned because coordination cost too much; the missing evidence is how many reach an acceptable delivery once we stop selecting only the demonstrations that went well.
Primary sources consulted on November 10, 2026. We have not run these models or reproduced the experiments. Adoption criteria and expectations are our own analysis.