Claude Fable 5.1 and Mythos 5.1

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026. They are the same underlying model. What separates them is the level of safeguard applied and who is permitted to run it: Fable 5.1 is generally available, and Mythos 5.1 is reachable only through trusted access programmes aimed at cybersecurity and the life sciences, restricted to vetted organisations and individuals in the United States.

The capability cases given for the restricted version are specific rather than general. In cybersecurity the newer safeguards refuse 60 percent fewer false positives, and the model is permitted to find software vulnerabilities while remaining refused on writing exploits for them. In the life sciences, Mythos 5.1 designed protein binders whose measured binding affinity on three targets was ten times higher than the best entries submitted to Adaptyv Bio's design competitions.

The pricing moved in a way that the headline figures hide. The base rates for input and output did not change. What changed is the cost of reading from cache, which fell 75 percent to $0.25 per million tokens from $1, making the model roughly 25 percent cheaper for ordinary work and up to 45 percent cheaper for the highly agentic kind.

Until now, a laboratory deciding a capability was too dangerous to release had two options: ship it to everyone or ship it to nobody. This is a third. The same weights exist in two configurations that differ only in what they will refuse, and access to the permissive one is a contractual relationship rather than a purchase. Whether that is a proportionate answer to dual-use capability or the beginning of a tiered internet for models is a question this museum is too close in time to settle, but the mechanism is now in use and is worth recording at the moment it appeared. Two days later OpenAI released GPT-6 Astra the same way, holding its cybersecurity capabilities back from general availability, which makes this a week rather than a company.

The pricing change is the quieter half and may matter more. An agent that works for hours re-reads the same context hundreds of times, so the cost of a long-running task is dominated by cache reads rather than by the tokens a person sees. Cutting that rate by three quarters while holding the visible price steady does not make the model look cheaper on a comparison table; it makes a category of work affordable that previously was not. The era this room describes is named for autonomy, and autonomy has always been priced by the hour.