Anthropic Launches Claude Fable 5.2 and Mythos 5.2, With Mythos Gated to Verified Organizations
Anthropic announced Claude Fable 5.2 and Claude Mythos 5.2 on 9 November 2026. Fable 5.2 launched on the day of the announcement, while Mythos 5.2 was initially available only to vetted organizations. The Verge reported that Anthropic credits both models with improvements in coding, reasoning and vision, and that both generate outputs faster than the 5.1 versions released in September.
Anthropic states that Fable and Mythos are the same underlying model. The difference lies in the safeguards layered on top. Fable 5.2 includes cybersecurity and biology safeguards that allow it to be made available to all users. Organizations that want to use Mythos for cybersecurity or biomedical research can apply through Anthropic's Life Sciences and Cyber Verification Programs.
On throughput, Anthropic says Fable 5.2 and Mythos 5.2 generate outputs up to 30% faster than Fable 5.1 and Mythos 5.1. Both models also offer an optional fast mode that speeds token generation by up to 2.5 times. Standard pricing is $8 per million input tokens and $40 per million output tokens. Fast mode is priced at $16 per million input tokens and $80 per million output tokens. On Medium effort, the default in Claude apps and Claude Code, Anthropic says Fable 5.2 completes challenging tasks with fewer tokens than Fable 5.1.
Anthropic also reports results from its FrontierCode 2.0 evaluation. Fable 5.2 at High effort scored about level with Opus 5.5 at Max effort, at roughly half the cost per task. Anthropic's announcement also describes prerelease testing in which Fable 5.2 and Mythos 5.2 formalized the proof of Fermat's Last Theorem in 30 hours, designed three working electronic devices from scratch, and solved two open problems in the physics of black holes. These are vendor-reported results from testing, and the company has not published the methodology behind them.
On safeguards, Anthropic says Fable 5.2's cyber and biology protections are as robust as those in Fable 5.1 but produce fewer unnecessary blocks. The company also says it implemented its strongest-ever protections against distillation attacks in both 5.2 models.
The 5.2 release follows a sequence of launches earlier in the year. In June, Anthropic launched Claude Fable 5 as a Mythos-class model made safe for general use. At that time the company described it as a step forward on agentic coding and prototyping, and said it works at senior research scientist grade, picking directions. A Reuters report dated 26 June said the U.S. government had allowed Anthropic to release Mythos to trusted U.S. companies, and that it was moving toward allowing Fable to be released, though the timeline was unclear. Mashable separately reported that Anthropic pulled Claude Fable 5 and Claude Mythos 5 after a Trump administration order barring foreign use. Anthropic's platform documentation says Mythos 5 is offered separately for defensive cybersecurity workflows and is available only to organizations verified through its verification process.
The broader context here is a tiering model. A single set of weights is being sold in two configurations, one open to all users under a set of safeguards and one gated behind organizational verification for sensitive domains. For security teams and life-sciences groups, the practical question is whether the verification programs are quick enough to matter for their work. For everyone else, the question is how far the 5.2 safeguards can narrow false positives without loosening the protections that justify general availability. Anthropic's claim of fewer unnecessary blocks is the number to watch in production, because refusal rates on legitimate security and biology queries are where this kind of gating tends to hurt most.
The cost and speed figures deserve the same scrutiny. A 30% throughput gain and a fast mode at 2.5 times the token rate, priced at double the standard rate, give engineers a clear trade between latency and spend. The FrontierCode comparison, if it holds across workloads, suggests that the cost per completed task may matter more than the cost per token for agentic coding. Teams running long, multi-step jobs will need their own measurements, since vendor evaluations seldom reproduce the exact mix of tools, retries and repository sizes they face.
What this release enables is clearer than what it costs. A model that can be used by any developer, with a verified tier for high-risk work and a fast mode for time-sensitive tasks, widens the set of problems that can be handed to an automated system. Whether the reported results on formal proofs, hardware design and black hole physics translate into routine research tooling is a question the next several months of customer use will answer.
