Anthropic Launches Claude Fable 5.1 With 75% Cheaper Caching and Fewer False Positives

News Summary

Anthropic on Tuesday, September 1, 2026 (Pacific Time) released Claude Fable 5.1 alongside a companion model, Claude Mythos 5.1, marking the first update to the Fable line since Claude Fable 5 launched on June 9, 2026. The new model is positioned as Anthropic's most advanced offering for coding, knowledge work, and long-running agentic problem-solving, and it ships with a notably lower price for cached and agentic workloads. Fable 5.1 is generally available to anyone with a Claude account starting today, while Mythos 5.1 — the same underlying model with different safety configurations — remains limited to trusted-access programs focused on cybersecurity and life-sciences research.

What's New in Fable 5.1

Anthropic says Fable 5.1 outperforms its predecessor, Claude Fable 5, as well as the company's Opus-tier flagship, on a broad set of coding and reasoning benchmarks. According to figures cited by outlets covering the launch, the model reaches roughly 55.8% on the Terminal-Bench coding evaluation, and it shows the largest gains over Fable 5 when run at higher "effort" settings — the parameter that controls how much the model reasons before answering. At low and medium effort, Fable 5.1 performs roughly on par with Fable 5, but Anthropic reports a much wider gap in its favor as effort increases, suggesting the model is tuned to make better use of extended reasoning on hard, multi-step tasks.

Anthropic frames the release around two practical themes for engineering teams: fewer wasted shortcuts and better root-cause fixes. The company says Fable 5.1 is less prone to taking shortcuts that produce lower-quality code, and more capable of tracing a bug back to its actual source rather than patching symptoms — a distinction that matters for teams running the model autonomously across long coding sessions.

Pricing Changes

Fable 5.1 keeps the same headline price as Fable 5 — $10 per million input tokens and $50 per million output tokens — but Anthropic cut the price of cache-read tokens sharply, from $1 to $0.25 per million tokens, a 75% reduction. Because agentic coding workflows often re-read large amounts of cached context (file trees, prior conversation history, tool outputs) on every turn, Anthropic estimates the effective cost of typical workloads drops by about 25%, and by as much as 45% for the most cache-heavy, highly agentic use cases. For engineering teams running long autonomous coding sessions, this cache-pricing change is likely to matter more in practice than the headline per-token rate.

Safety Filter Adjustments

Alongside the capability and pricing updates, Anthropic revised the safety classifiers that sit in front of Fable-class models. The company reports that cybersecurity-related false positives — cases where a legitimate, benign request was incorrectly flagged and declined — have dropped by roughly 60% compared with Fable 5. On biology and medicine topics, the reduction is even larger: Anthropic cites an 85% drop in filters triggering on harmless questions about basic biology and medical topics.

The changes also extend what the generally available Fable 5.1 model is permitted to do in security contexts. The model can now be used to help discover software vulnerabilities, a task that previously required more restrictive handling. Anthropic draws a clear line, however, around exploit development: three categories of dual-use security work — penetration testing, exploit generation, and binary-based vulnerability scanning — are still redirected away from Fable 5.1 toward Anthropic's Opus-tier models, which carry a different safeguard configuration.

Fable vs. Mythos

Claude Fable 5.1 and Claude Mythos 5.1 are, per Anthropic, the same underlying model differentiated only by their safeguard settings. Fable 5.1 is available broadly to Claude account holders worldwide starting with today's launch. Mythos 5.1 is reserved for participants in Anthropic's trusted-access programs — vetted organizations and individuals, described as limited to U.S. companies and qualified individuals, working specifically in cybersecurity research and life-sciences fields where the standard safety posture would otherwise be too restrictive for legitimate research.

Why It Matters

The Fable 5.1 release illustrates a pattern that has become common among frontier AI labs: rather than waiting for a full next-generation model, Anthropic is iterating on safety-filter precision, cost efficiency, and benchmark performance within the same model generation. For developers and enterprises already building on Claude Fable 5, the update is designed as a low-friction upgrade — same top-line pricing, meaningfully cheaper effective cost for cache-heavy agentic pipelines, and a model tuned to waste less effort on shortcuts during long coding runs. For the broader public conversation about AI and safety, the sharp drop in false-positive refusals on both cybersecurity and biology topics signals Anthropic's continued effort to narrow the gap between blocking genuinely harmful requests and frustrating legitimate researchers, students, and engineers with overly cautious filters.