Short answer: It took an unreleased research version of Claude 11 days to formalize the proof of Fermat's Last Theorem. On November 9, Anthropic said a team of 1,000 Claude Fable 5.2 agents did it in 30 hours, with minimal human help. The same day, the price of the model dropped by 20%.
That is the headline. The details are more interesting: where Fable 5.2 actually jumped, what Anthropic still keeps behind a door, which of your requests may be answered by a different model, and why the real discount for agent work is bigger than 20%.
What did Anthropic ship on November 9?
Two models. Claude Fable 5.2 (API ID claude-fable-5-2) is open to every developer on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry. Claude Mythos 5.2 has the same price but goes only to organizations verified through Anthropic's Cyber Verification Program and Life Sciences Verification Program.
Both keep a 1M-token context window, up to 128K output tokens and a June 2026 knowledge cutoff. Anthropic says both generate output up to 30% faster than the 5.1 models. Fable 5.1 does not disappear: it is now a legacy model, with retirement no sooner than September 1, 2027.
The release landed shortly before Anthropic's planned IPO, which is how CNBC framed it. For developers, the timing matters less than what changed in the model and in the bill.
How did 1,000 agents formalize Fermat's Last Theorem in 30 hours?
The proof was written in Lean, a programming language for mathematics where every step has to check. According to Anthropic, Fable 5.2 started from minimal prompts, built its own ways to coordinate, and used no mathematical references, only a list of intermediate theorem statements to score progress.
About 1,000 Fable 5.2 agents formalized the proof of Fermat's Last Theorem in Lean in 30 hours, according to Anthropic. A single run, not a benchmark.
The comparison is the striking part. Opus 5.5, run with the same setup of about 1,000 agents, completed only 26% of the formalization in 18 hours and, in Anthropic's words, barely made any more progress after 30. Both lines come from a single run per model, so treat it as a demonstration, not a benchmark.
The lesson for anyone building agents: the gain shows up when a model has to keep a long, shared plan alive across many workers. That is exactly where the previous models stalled.
Where did Fable 5.2 actually get better?
Not evenly. On the coding benchmarks most people quote, the jump is solid but not dramatic: FrontierSWE v2 went from 56.3% on Fable 5.1 to 69.9%, and OSWorld 2.1 from 83.5% to 89.3%. Humanity's Last Exam with tools rose from 65.6% to 74.2%.
The big moves are in long, messy, real work. On the Remote Labor Index, Fable 5.2 scored 41.3% at high effort against 17.9% for Fable 5.1, more than double. On RoboDojo, where a model controls a simulated two-arm robot from camera images, it scored 64.9% against 27.6% for Fable 5.1 and 29.8% for GPT-6 Astra.
All of these numbers are reported by Anthropic. They still tell a clear story: Fable 5.2 is less a better autocomplete and more a better worker on tasks that run for hours.
What is Mythos 5.2, and why can't you use it?
Mythos 5.2 is the same generation built for dual-use domains: security research and biology. Anthropic says it found 10.2% more vulnerabilities than Opus 5.5 and 14.4% more than Mythos 5.1, and it scored 89.6% on CVE Rediscovery against 78.3% for Mythos 5.1.
Claude Mythos 5.2 is available only to organizations verified through Anthropic's Cyber and Life Sciences programs.
That capability is the reason for the door. Access requires verification, and even fast mode for Mythos 5.2 is limited to members of the verification programs. Claude Enterprise customers can reach it through Claude Security. Everyone else gets Fable 5.2.
Which of your requests get answered by another model?
Here is the detail most launch coverage skipped. Fable 5.2 runs safety classifiers on cybersecurity and biology topics. When a query is blocked, Anthropic says the user receives a response from Opus 4.8 for cybersecurity and from Opus 5 for biology. A small set of AI research topics also falls back to Opus 5.
When a classifier blocks a query, Anthropic says the answer comes from Opus 4.8 (cybersecurity) or Opus 5 (biology).
Anthropic calls these fallbacks transparent and reports about 55% fewer blocks than Fable 5.1 in early testing across claude.ai, Claude Code and Cowork. If you build security or bio tooling on the API, read Anthropic's refusals and fallback guide before you assume every answer came from Fable 5.2.
Two more fine-print items. Fable 5.2 requires 30-day data retention and is not available under zero data retention unless Anthropic expressly authorizes it. And fast mode, which Anthropic says speeds generation up to 2.5 times, costs twice the standard price.
How much cheaper is it, really?
On the price list, exactly 20%. Fable 5.2 costs $8 per 1M input tokens and $40 per 1M output tokens, against $10 and $50 for Fable 5.1. Cache reads cost $0.20 per 1M tokens, the Batch API halves the price to $4 and $20, and there is no surcharge for long prompts anywhere in the 1M window.
Fable 5.2 lists at $8 input and $40 output per 1M tokens, 20% below Fable 5.1.
For agents, the cut is deeper because of what we call the Cache Gap. GPT-6 Astra lists at $10 and $50, only 20% above Fable 5.2, but charges $1.00 per 1M cached tokens, five times more. Agent tools resend the same context every turn, so most input is cached. In our calculation of a cache-heavy coding session, Fable 5.2 costs $17.44 and Astra $27.20, a 35.9% difference. The full breakdown is in our Claude Fable 5.2 pricing study.
It is still a premium model. The same session costs $9.44 on Claude Opus 5.5, so Fable 5.2 runs 1.85 times the price. On Tokely, Fable 5.2 lists at $6 input and $30 output per 1M tokens.
Should you switch?
If you run long agent jobs, multi-step research or anything that has to hold a plan for hours, yes. That is where the Remote Labor Index and the Fermat run say Fable 5.2 pulls away, and it costs less than the model it replaces.
If most of your traffic is short chat turns or routine code edits, Opus 5.5 or Sonnet 5.5 will do the job for a fraction of the cost. A practical setup routes everyday turns to the cheaper model and sends only the hardest tasks to Fable 5.2. Keep fast mode for the moments when a person is waiting.
If you work in security or biology, test your real prompts first. Some answers may come from Opus 4.8 or Opus 5, and that changes what you are actually buying.
Originally published at tokely.me.
Top comments (1)
The interesting part isn't just the price reduction or the number of agents. I'd be more interested in how these changes affect reliability, coordination, and the actual cost of running multi-agent workflows in production. A larger agent count is useful only if it improves results without adding unnecessary complexity.