Fable 5.1 and Mythos 5.1 have just been released.
It tops all eight public benchmarks, with prices cut by up to 45%.
Notably in scientific research and coding, it clearly outpaces Fable 5 and rival GPT-5.6 Sol.
To put it more intuitively: running Fable 5.1 with medium or low reasoning effort delivers performance roughly equivalent to the older Mythos 5 at very high or even maximum reasoning effort.
This round, "Old Fable" has taken down "Mythos Saint."
On pricing, Fable 5.1 cuts cache-read prices to $0.25 per million tokens, a 75% reduction.
Input and output prices remain unchanged from Fable 5, at $10 and $50 per million tokens, respectively.
It may look like only one price was cut, but the impact is significant because cache reads account for the bulk of costs in agentic tasks.
Anthropic says costs for typical workloads are about 25% lower than Fable 5, and for highly agentic tasks, savings can reach up to 45%.
The official release video states:
This is possible because it excels at handling multi-step tasks.
In such tasks, a tiny error at step two can derail the entire process by step 40, but Fable 5.1 maintains stable output throughout.
If it encounters an unsolvable problem, it will tell you which approaches it has tried and where it got stuck.
Researcher Flix Rieseberg also noted that in writing, Fable 5.1 uses less bold text, fewer headings, lists, or quotes, and follows style prompts more closely.
That said, could you also open-source your style prompts?
As usual, Fable 5.1 is available to all users, while Mythos 5.1 is only offered through a trusted access program to vetted cybersecurity and life sciences institutions.
Research in action: protein design, generating a Venus map, GPU acceleration.
Beyond benchmarks, Anthropic also demonstrated the real-world research results of Fable 5.1 and Mythos 5.1.
For protein design, Anthropic had Mythos 5.1 use open-source protein design and folding tools to design high-affinity binding proteins, and sent the designs to two external organizations for experimental validation.
On three targets, the binding affinity of Mythos 5.1's designs was 10 times that of the best solution in the Adaptyv Bio protein design competition. Across all 12 targets, the hit rate approached 50%, compared with the typical 10% to 15% in current protein design.
High-affinity binding proteins are the first step in many common drug development pipelines, directly determining whether a drug can work at lower doses.
In astronomy, Fable 5.1 trained a neural network on radar images captured by NASA's Magellan spacecraft more than 30 years ago, generating new high-resolution topographic maps for one-third of Venus's surface.
Previous topographic maps covered only one-fifth of Venus at resolutions between 10 and 20 kilometers. Fable 5.1 improved resolution to 2 to 3 kilometers and increased height accuracy by 25% over previous maps. The map has been released under a Creative Commons license for the upcoming NASA VERITAS and ESA EnVision missions, helping to identify key geological features for future observations.
In computational biology, Mythos 5.1 wrote custom GPU kernels and cached intermediate results to speed up seven open-source deep learning models by up to 2.5 times, with identical outputs.
In real-world research, biologists may need to run these models thousands of times to test every possible mutation near each human gene, and the optimized model can cut GPU costs by 30% to 60%.
Such optimization typically takes a performance engineering team weeks to complete, putting it out of reach for many academic labs. Mythos 5.1 accomplished it in just days, relying solely on public source code.
Anthropic plans to open-source these optimizations in the near future.
The Fable 5.1 release also comes with special provisions and API changes.
First, a brief note on safety.
Fable 5.1's network-security false-positive rate is 60% lower than Fable 5's. It is now permitted for discovering software vulnerabilities, but developing exploit code remains prohibited. Dual-use tasks such as penetration testing, exploit generation, and binary-based vulnerability scanning are still redirected to the Opus series models.
The false-positive rate for basic biology and medical questions fell by 85%, but queries involving life-sciences R&D are still routed to Opus models. Professionals must obtain access through Mythos 5.1's Life Sciences Validation Program (LSVP).
Next, the headline addition: an anti-distillation mechanism.
Starting today, newly registered API accounts can no longer manually edit Claude's context in multi-turn conversations while preserving chain-of-thought records.
Previously, a commonly documented technique involved manually altering Claude's earlier context in a multi-turn conversation while deliberately retaining the chain-of-thought records. This allowed replaying the model's reasoning generated under one set of instructions under a new, potentially adversarial set of instructions.
That path is now completely closed.
The fix is to attach a signature to each chain-of-thought block.
When a user sends an assistant reply back to the API in a subsequent request, the system uses the signature to do two things: verify that the current model is authorized to read the block, and verify that the entire conversation preceding the block—including the system prompt, tool list, and all historical messages—is identical to what it was when the block was originally generated.
If anything in the conversation has been altered, the signature check fails.
The handling of a failure depends on a developer-set parameter, prefix_mismatch_behavior. The default value is "error," which returns a 400 status code and rejects the request. If set to "drop_block," the system silently discards the block and all subsequent chain-of-thought content, while the request itself proceeds normally.
The discarded content is not billed—a nice touch—and is listed in the input_transformations array in the response.
Which operations trigger validation failure?
The documentation provides a detailed list: editing, rearranging, or deleting any prior user/assistant/system messages; altering the top-level system prompt; adding, removing, or renaming any tools in the tool list; removing a chain-of-thought block from the middle of the conversation history while keeping later blocks; and referencing an image or document URL that returned different content between two requests.
Any one of these will invalidate all subsequent chain-of-thought blocks.
Conversely, some operations are safe:
Appending new messages at the end of the conversation, removing chain-of-thought blocks from the beginning of the history, changing request parameters such as max_tokens, adding or removing cache_control markers, and server-side compaction or context editing will not trigger validation failure.
There is also a chained validation design: each chain-of-thought block records information about its predecessor, forming a chain across turns. Blocks can be removed from the head of the chain, but if one is removed from the middle, every block after it becomes invalid, and the entire chain from the break point onward is void.
To avoid trapping developers, Anthropic also offers a range of alternatives that achieve the same effect without touching the conversation history.
Need to change instructions mid-conversation? Use a mid-conversation system message instead of directly editing the top-level system prompt.
Need a temporary reminder each turn? Use a turn-level system message with a clear_at parameter, rather than the old insert-then-delete approach.
Need to add or remove tools mid-conversation? Use tool_addition and tool_removal blocks, rather than directly editing the tool list.
Need to trim context? Use server-side compaction and context editing, rather than crude client-side truncation.
If you're using official products like Claude Code, claude.ai, Claude Managed Agents, or the Claude Agent SDK, nothing needs to change—they already automatically ensure the conversation prefix can't be tampered with.
But if you're a developer calling the Messages API directly, Anthropic recommends treating the messages array as strictly append-only, then running a full multi-turn conversation test with prefix_mismatch_behavior set to 'drop_block' to see if any prefix_binding_mismatch entries appear in the logs.
The documentation also highlights a particularly tricky scenario.
If you maintain a tool or framework where users bring their own API keys, newly registered users will hit this validation sooner. That's because developers' own keys were likely registered before August 31 and aren't yet subject to enforcement.
In that case, it's best to proactively set prefix_mismatch_behavior and test ahead of time, rather than waiting for users to hit errors before you discover the issue.
One More Thing
Anthropic's big move is sure to get a reaction from OpenAI.
However, OpenAI's new model Astra is not quite ready, so it only issued a symbolic preview.
Ilya also made a rare public appeal for stronger cybersecurity as Fable 5.1 and OpenAI Astra launch.
Reference link: [1] https://www.anthropic.com/claude-fable-and-mythos-5-1