Anthropic shipped Claude Fable 5.1 and Claude Mythos 5.1 this week, three months after the Fable 5 line landed. The two releases are the same underlying model wearing different safeguard layers. Fable 5.1 is generally available as claude-fable-5-1 across the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry. Mythos 5.1 stays locked inside Project Glasswing for vetted US organizations. Both carry a 1M token context window, 128K max output, and adaptive thinking that never turns off.
The headline number is hard to ignore: 52.6% on Terminal-Bench-Science 0.1, a scientific research benchmark. That’s roughly double Fable 5’s 24.7% and well ahead of Opus 5’s 29.0% and GPT-5.6 Sol’s 22.4%. Anthropic reports a standard error of 3.5 to 4.5 points per model, so the exact ranking matters less than the trajectory. On Terminal-Bench 4.0, Fable 5.1 hits 55.8% while Mythos 5.1 reaches 60.9% — and the gap between those identical models is the cost of safeguard interventions. That’s an unusually honest disclosure in a field where safety overhead is rarely quantified.
Pricing is where commercial teams will feel the change. Base input and output rates stay at $10 and $50 per million tokens, but cache reads drop 75% from $1.00 to $0.25 per million. Anthropic estimates roughly 25% lower cost on typical workloads and up to 45% on context-heavy agentic ones. For teams running long-horizon agents that hammer the same context repeatedly, that’s real money.
But there are breaking changes that will bite. Forced tool use is gone — setting tool_choice to “any” or “tool” now returns a 400. Thinking blocks are model-bound, meaning earlier models can’t read Fable 5.1’s reasoning, so router and fallback setups lose that signal when they switch down. Editing earlier turns invalidates thinking blocks entirely. The enforcement date is August 31, 2026 for new accounts. Anthropic also documents regressions: parallel tool calling is more variable, the model narrates less, and it prefers whole-file rewrites over targeted edits. The science results are genuinely impressive — Mythos 5.1 designed protein binders at roughly 50% hit rate against a 10-15% norm — but teams should test their agent loops before assuming a drop-in upgrade.
