Anthropic releases Fable 5.1 and Mythos 5.1, topping benchmarks and cutting costs

Fable 5.1 is available to all users, while Mythos 5.1 stays limited to approved cybersecurity and life-science organizations under the trusted access program. The report said Fable 5.1 clearly outperformed Fable 5 and OpenAI's GPT-5.6 Sol on scientific research and coding. Running Fable 5.1 at medium or low reasoning effort roughly matches the previous Mythos 5 at extremely high or maximum reasoning effort, according to QbitAI.

Prices for Fable 5.1 remain the same as Fable 5 for input and output, at $10 and $50 per million tokens, but cached reads are cut to $0.25 per million tokens, a 75% reduction. Since cached reads account for a large share of agentic workload costs, the price cut produces the overall savings. Anthropic said Fable 5.1 is better at multi-step tasks: a small error at step two can otherwise crash the entire process by step 40. When the model gets stuck, it reports which approaches were tried and where the bottleneck is. Researcher Flix Rieseberg said Fable 5.1 uses less bold text, fewer headings, lists and quotes in writing, and follows style prompts more closely.

Anthropic also showed scientific results. Using open-source protein design and folding tools, Mythos 5.1 designed high-affinity binding proteins that were tested at two external labs. On three targets, the proteins achieved binding affinities ten times those of the best entries in the Adaptyv Bio protein design competition; on all twelve targets, the hit rate approached 50%, compared with the typical 10-15% in protein design. For astronomy, Fable 5.1 trained a neural network on radar images captured by NASA's Magellan spacecraft more than 30 years ago, generating a new high-resolution terrain map for one-third of Venus. The resolution was improved to two to three kilometers, and height accuracy was 25% better than earlier maps that covered only one-fifth of the planet. The map is available under a CC license for future NASA VERITAS and ESA EnVision missions.

In computational biology, Mythos 5.1 accelerated seven open-source deep learning models by up to 2.5 times by writing custom GPU kernels and caching intermediate results, while keeping outputs exactly the same. Anthropic said this sort of optimization typically consumes weeks for a performance engineering team and is unaffordable for many academic labs; Mythos 5.1 accomplished it in days using only public code. Anthropic plans to open-source the optimizations soon.

Alongside the release, Anthropic introduced a new anti-distillation mechanism. The report said Fable 5.1's false-positive rate for cyber security fell 60% and the model can now be used to find software vulnerabilities, but not to develop exploits. Dual-use tasks such as penetration testing, exploit generation, and binary-based vulnerability scanning are still redirected to Opus models. For basic biology and medical questions, the false-positive rate fell 85%; queries related to life-science R&D are still routed to Opus, with professional access through Mythos 5.1's Life Sciences Verification Program.

The anti-distillation mechanism adds a signature to every chain-of-thought block. When assistant replies are returned to the API, the signature is used to check both whether the current model is allowed to read the block and whether the full preceding conversation prefix has not changed. Editing, reordering, or deleting earlier messages; changing the top-level system prompt; altering the tool list; or extracting an intermediate chain-of-thought block while keeping later blocks will invalidate all subsequent blocks. A chained check links each block to the one before it, so removing one from the middle invalidates every block after the break. Developers can set prefix_mismatch_behavior to error, producing a 400 response, or to drop_block, which silently discards the mismatched block and later chain-of-thought; dropped content is not billed. Appending messages, removing blocks from the start, changing request parameters, and server-side compression are considered safe, Anthropic said.

Anthropic said official products like Claude Code, claude.ai, Claude Managed Agents, and the Claude Agent SDK already preserve prefix integrity automatically. For direct Messages API developers, the company recommends treating the messages array as append-only and testing in drop_block mode to watch for prefix_binding_mismatch entries. It also lists alternatives such as mid-conversation system messages, per-turn system messages with a clear_at parameter, tool_addition and tool_removal blocks, and server-side context editing to avoid breaking history.

The report also said that OpenAI's Astra model was not fully ready around the same time and only received a preliminary announcement. Ilya made a rare call for stronger cyber security as Fable 5.1 and Mythos 5.1 were released.