Claude Fable 5.1: 52.6% Science Benchmark Score

Claude Fable 5.1 is Anthropic's generally available model for longer-running coding and knowledge work, while Claude Mythos 5.1 uses the same underlying model behind more permissive safeguards for vetted organizations. Anthropic reports a 52.6% score for Fable 5.1 on Terminal-Bench-Science 0.1, a 75% reduction in cache-read pricing, and availability through the Claude API and major cloud platforms. The release matters most to teams running agents that repeatedly reuse large contexts, a practical extension of the agent execution patterns covered elsewhere on Yowox and the frontier coding context already relevant to Claude users.

Definition: Claude Fable 5.1 is Anthropic's latest generally available model for agentic coding, research and knowledge work.

Example: A coding agent can reuse repository context at $0.25 per million cache-read tokens instead of $1.00.

Key takeaway: Fable 5.1 combines a reported 52.6% science-benchmark score with lower context-reuse costs, but migration still requires compatibility testing.

Business impact: Teams with long-running agents may see lower token bills and fewer model escalations, while teams with short prompts may see less benefit.

What Anthropic released

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 as two safeguard variants of one underlying model on September 1, 2026. Claude Fable 5.1 is generally available with the model identifier claude-fable-5-1, while Claude Mythos 5.1 is limited to vetted organizations using trusted access programs. The official Anthropic announcement is the reference for the release's benchmark, pricing and access claims.

Claude Fable 5.1 has a 1 million token context window, a 128,000-token maximum output, and adaptive thinking. Claude Fable 5.1 is available through the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry, according to the launch coverage. Operators should verify the exact account, region and platform terms before treating availability as a production commitment.

How much faster is Fable 5.1 on the reported benchmarks?

Claude Fable 5.1 reaches 52.6% on Terminal-Bench-Science 0.1 in Anthropic's reported evaluation, compared with 24.7% for Claude Fable 5, 29.0% for Claude Opus 5 and 22.4% for GPT-5.6 Sol. Anthropic also reports a 3.5–4.5-point standard error per model, so teams should reproduce the relevant workload rather than infer a universal capability ranking from one result.

Claude Fable 5.1's benchmark profile is strongest on long-horizon and agentic work, while the CursorBench increase is smaller at 73.4% versus 70.5% for Fable 5. The practical takeaway is to compare complete task trajectories—tool choices, retries, token usage and review burden—rather than compare isolated answer quality.

Why the cache-read price cut matters

Claude Fable 5.1 cache reads now cost $0.25 per million tokens instead of $1.00, a 75% reduction that applies where usage is billed by token. Anthropic says the reduction translates to about 25% lower total cost on typical workloads and up to approximately 45% on highly agentic workloads; base input remains $10 per million tokens and output remains $50. The source coverage describes the same distinction between cache pricing and unchanged base rates.

Cache reads matter when an agent repeatedly sends already-processed instructions, repository context or tool results. Claude Fable 5.1 therefore has a stronger economic case for coding agents and other context-heavy systems than for short, one-shot prompts. Teams should measure cache-hit share and blended input/output spend, because a 75% cut to one line item does not make every workload 75% cheaper.

What changes for API migrations?

Claude Fable 5.1 requires migration testing around forced tool use, thinking blocks and edited conversation history. The release coverage reports that tool_choice values of any or tool return a 400, earlier models cannot read Fable 5.1 thinking blocks, and modifying earlier turns or rebuilding system and tool arrays can error for some newer accounts. Teams should switch to supported tool-selection and context-editing patterns before routing production traffic.

Claude Fable 5.1 also changes operational behavior beyond request validation. Anthropic's release notes describe more variable parallel tool calling, less narration, more memory-based answers at low effort, and a preference for whole-file rewrites over targeted edits. A migration test should therefore record tool-call counts, diff scope, refusal or fallback behavior, latency and the quality of the final artifact—not just whether the request returns HTTP 200.

Who gets Claude Mythos 5.1?

Claude Mythos 5.1 is the same underlying model as Claude Fable 5.1 but is restricted to vetted cybersecurity and life-sciences organizations through trusted access programs. Anthropic says Mythos supports more permissive safeguards for eligible defensive-security and professional research work while retaining other safeguards. Teams cannot treat Mythos 5.1 as a generally available alternative endpoint.

Claude Mythos 5.1's access model explains why its Terminal-Bench 4.0 score of 60.9% should not be read as an ordinary public API comparison with Fable 5.1's 55.8%. The difference reflects safeguard interventions and access conditions, so an operator comparing the models should document which model, safeguards, tools and task policies were active.

What the science examples show

Claude Mythos 5.1 designed protein binders that Anthropic says reached nearly 50% viability across 12 targets, compared with a stated 10–15% typical hit rate, while Claude Fable 5.1 produced a higher-resolution elevation map for part of Venus from older radar data. These are Anthropic-reported research examples, not independent replication, so research teams should treat them as demonstrations of possible workflow value rather than guaranteed scientific outcomes.

Claude Mythos 5.1 also wrote custom GPU kernels that Anthropic says sped up seven open-source genomics models by up to 2.5 times with identical outputs. For operators, the concrete lesson is to validate generated kernels, numerical equivalence and end-to-end cost on the target hardware before counting a reported speed-up as a production saving.

What teams should test next

Teams evaluating Claude Fable 5.1 should start with a fixed sample of real agent tasks and compare the current model against claude-fable-5-1 under the same tools, context, effort limits and acceptance checks. The benchmark signal and cache pricing make Fable 5.1 a credible candidate for long-running work, but safeguards, tool-call patterns and edit behavior can change the total cost of a successful task.

The next useful signal is not another headline score; it is whether Claude Fable 5.1 completes the team's work with fewer retries, acceptable review effort and a lower blended cost. Teams should keep a fallback route to another model until those measurements hold across representative tasks.

FAQ

The release is substantial, but its value is workload-dependent. Claude Fable 5.1's benchmark gains and cache economics favor agents that work through large, repeated contexts; Mythos 5.1's access restrictions keep its most permissive behavior out of the general developer path. The safest decision is a measured migration pilot with explicit compatibility and quality checks.