What happened
Anthropic’s November 9, 2026 announcement describes Fable 5.2 and Mythos 5.2 as the same underlying model, with different safeguards and access. Fable has cybersecurity and biology safeguards and is available to all users. Organizations working on beneficial cybersecurity or biomedical research can apply for Mythos through Anthropic’s Cyber and Life Sciences Verification Programs. The company says the new models communicate more clearly than earlier versions and generate outputs up to 30% faster than Fable 5.1 and Mythos 5.1.
The listed price is $8 per million input tokens and $40 per million output tokens. An optional fast mode, which Anthropic says can increase token-generation speed by up to 2.5 times, costs $16 per million input tokens and $80 per million output tokens. Anthropic recommends the models for ambitious, open-ended technical work, while suggesting that users start with Claude Opus 5.5 for most coding and knowledge tasks. It also describes Fable as a possible Advisor for smaller models or Orchestrator that delegates simpler work.
The announcement reports gains across coding, search, computer use, science, and other evaluations. On FrontierCode 2.0, Anthropic reports a 49.1% score for Fable 5.2 at Xhigh effort, compared with 32.4% for Fable 5.1 and 45.5% for Opus 5.5. The company says that at High effort Fable 5.2 scored 45.3%, close to Opus 5.5’s 45.5% best score, for roughly half the cost per task. These are Anthropic’s reported evaluation results; effort settings differ across some comparisons.
The mathematical result concerns formalization: translating a proof into Lean, a language in which a computer can check each step. Anthropic says 1,000 Fable 5.2 agents completed the formalization in 30 hours with minimal human intervention, starting from minimal prompts, devising coordination methods, and using a list of intermediate theorem statements to score progress. The resulting proof was seven times shorter and used one-third as many action tokens as the company’s earlier formalization, which took 11 days and involved a purpose-built harness and human guidance. In a comparison using the same 1,000-agent setup, Anthropic says Opus 5.5 completed 26% after 18 hours and 27% after 30 hours. Each model’s result came from a single run. The claim is about formalizing Fermat’s Last Theorem, not discovering a new proof of it.
What the archive adds
New Runtime’s earlier coverage of AlphaProof Nexus solving nine open Erdős problems described formal proof search as valuable when candidate results are tied to machine-checkable obligations. This is a new event on that thread: Anthropic reports a large coordinated model team completing a formalization, with progress scored against intermediate theorem statements.
Trend signal
This supports the view that scientific-agent progress is most useful when it leaves inspectable, checkable artifacts. A Lean formalization is a concrete verification target rather than a model’s claim that its reasoning is correct. The relevant New Runtime pattern is scientific agents need artifact evidence; the announcement’s single-run comparison, however, is not independent evidence of repeatability.
Why now
Anthropic presents the release as both a model upgrade and a change in the economics of using its higher-capability models: the listed standard token prices are lower than those for the preceding versions, and the company reports faster generation. That makes the announcement relevant to teams reconsidering when to route difficult work to a more capable model rather than a daily-driver model.
Connections to other events
The Fermat formalization connects directly to New Runtime’s earlier coverage of AlphaProof Nexus, which also centered on formal mathematics and computer-checkable proof obligations. The present announcement adds a different operational claim: a large agent team built its own coordination approach and used intermediate statements to measure progress. The two reports concern distinct systems and results, so they should not be read as a direct benchmark comparison.
Why it matters
For teams building with AI, the useful shift is from treating a long answer as completion to defining an external condition that can establish completion. In this example, Lean checks proof steps, while intermediate theorem statements give the agents a progress signal. That is a more concrete design pattern than simply asking many agents to work for longer.
The economics may also change model-routing decisions. Anthropic says Fable 5.2 at High effort approached Opus 5.5’s best FrontierCode score at roughly half the cost per task, while the announced per-token price is lower than the prior generation. Those figures make it worth testing the model on a team’s own difficult tasks, but they do not establish that it will be cheaper for every workload: task success, retries, tool use, and review effort all affect total cost. The practical lesson is to evaluate completed work and verification cost alongside token prices.
Limits and uncertainty
The performance figures and use cases are reported by Anthropic, and the announcement does not establish independent replication of the headline Fermat result. The comparison with Opus 5.5 used one run per model, so variability across runs remains unknown. The announcement describes the proof as a formalization checked in Lean, but the result does not by itself show that the same workflow will generalize to other mathematical problems. Reported benchmark scores also reflect specified effort levels and, in some cases, internal evaluations; they should not be treated as direct measures of production outcomes.
Skill to build now
For one difficult task this week, define a machine-checkable completion test before assigning work to an agent team. Break the goal into intermediate statements or tests that reveal progress, and record retries, tool costs, and human review time. Compare the resulting verified artifact—not just the model’s final explanation—with your existing workflow.
Reader takeaway
Fable 5.2’s headline is not simply that a thousand agents can work for 30 hours. It is that the reported workflow had an external verification target and intermediate progress checks. If you are testing agent orchestration, borrow that design principle first; treat the speed, cost, and single-run result as claims to validate on your own workload.