Anthropic shipped Claude Fable 5.1 and a restricted-access variant called Claude Mythos 5.1 on September 1-2, 2026, with the largest science benchmark jump in the Fable series yet and a 75 percent reduction in cache-read pricing that cuts typical workload costs by roughly 25 percent. The two models share identical underlying weights but differ in safeguards; Fable 5.1 is available to all API customers today, while Mythos 5.1 is limited to vetted cybersecurity and life-sciences researchers through Anthropic's Project Glasswing verification programs.

What Changed in the Fable 5.1 Benchmarks?

The most striking gain appears in scientific research capability. On Terminal-Bench-Science 0.1, Fable 5.1 scored 52.6 percent, more than doubling Fable 5's prior score of 24.7 percent. This represents the single largest benchmark improvement across Anthropic's entire announcement and signals a deliberate focus on agentic scientific workflows rather than coding alone.

Coding benchmarks also improved, though less dramatically. Fable 5.1 reached 55.8 percent on Terminal-Bench 4.0, up from Fable 5's 42.0 percent, while the restricted Mythos 5.1 variant scored 60.9 percent on the same benchmark. The most trustworthy independent confirmation comes from Cursor, the AI-native code editor, which posted on launch day that Fable 5.1 scored 73.4 percent on CursorBench 3.2.0 at maximum effort, calling it "the most capable model we've run on CursorBench 3.2".

Other benchmarks showed more modest gains. On OSWorld 2.0, a computer-use benchmark, Fable 5.1 improved from 72.9 percent to 77.9 percent with partial credit and from 36.1 percent to 41.7 percent with strict scoring. AutomationBench, which measures business workflow automation, jumped from 17.1 percent to 31.4 percent, nearly doubling.

How Does Pricing Actually Change for Users?

The headline pricing for input and output tokens remains unchanged. Fable 5.1 costs $10 per million input tokens and $50 per million output tokens, identical to Fable 5. However, the real savings come from prompt caching, a feature that stores frequently reused context like system prompts or large codebases so they don't need to be reprocessed on every request.

Cache-read pricing dropped from $1.00 per million tokens to $0.25 per million tokens, a 75 percent reduction. For agentic workflows that repeatedly reference the same large context window, this compounds quickly. Anthropic estimates the net effect is roughly 25 percent cheaper for typical workloads and up to roughly 45 percent cheaper for highly agentic work that leans heavily on cache hits.

What Are the Key Differences Between Fable 5.1 and Mythos 5.1?

Both models use the same underlying weights and architecture, but Mythos 5.1 has select safeguards lifted for approved users. Fable 5.1 remains available to all customers and maintains full cybersecurity safeguards, meaning it can discover vulnerabilities defensively but will not help develop exploits. Mythos 5.1 is restricted to participants in two verification programs: the Cyber Verification Program for vetted cybersecurity defenders and the Life Sciences Verification Program, which enrolled its first participants through a U.S. government partnership.

The performance gap reflects this trade-off. Mythos 5.1 scores 60.9 percent on Terminal-Bench 4.0 coding tasks, compared to Fable 5.1's 55.8 percent, suggesting that reduced safeguards allow slightly higher performance on certain benchmarks. Anthropic's own Claude Security product now runs on Mythos 5.1 for codebase vulnerability scanning and patch suggestion.

How to Optimize Your Prompts for Fable 5.1

If you are already running Fable 5 in production, Anthropic's engineering team recommends several practical steps to get the most out of the 5.1 upgrade without rewriting your entire prompt setup:

• Re-run your effort sweep: Fable 5.1 at low effort is often competitive with Claude Opus and Claude Sonnet models on cost per task while scoring higher. On CursorBench 3.2.0, Fable 5.1 at low effort matches Fable 5 at high effort for roughly one-third of the cost, so your previous effort-level settings may no longer be optimal.
• Change effort mid-conversation without breaking cache: A new feature in 5.1 allows you to raise effort for difficult steps and drop it for routine ones within the same conversation, something that previously invalidated the prompt cache. This is particularly useful for agentic workflows that mix hard reasoning tasks with simpler tool calls.
• Leverage the improved writing style: Anthropic explicitly addressed verbose and jargon-heavy output, a persistent complaint about earlier Fable models. The 5.1 release includes fixes for this, so you may see cleaner, more direct responses without additional prompt engineering.

What Does This Mean for the Broader AI Market?

The Fable 5.1 release arrives as the AI model market continues fragmenting into specialized tiers. Rather than a single "best" model, teams now choose based on workload and budget. Fable-class models target cost-conscious teams running agentic workflows, while Opus-class models serve teams prioritizing raw capability over cost. The 75 percent cache-read discount makes Fable 5.1 particularly attractive for long-running sessions that repeatedly reference the same context, a common pattern in code generation and scientific research agents.

The science benchmark jump also signals Anthropic's strategic direction. While competitors like Google focus on coding and general reasoning, Anthropic is explicitly building Fable-class models for scientific research workflows, a narrower but potentially high-value market segment.

Anthropic also reset usage limits for all users at launch, removing previous 5-hour and weekly rate limits, which may indicate confidence in the model's stability and a push to encourage heavier adoption.