Anthropic Releases Claude Fable 5.1 with Cheaper Cache

Anthropic has released Claude Fable 5.1 and Mythos 5.1, rolling out significant updates to its flagship model family. While the base pricing on OpenRouter remains $10.00 per million input tokens and $50.00 per million output tokens, Anthropic has slashed cache read prices from $1 to $0.25 per million tokens. This change reduces typical per-token costs by 25% and cuts agentic workflow expenses by up to 45%, including a 38% reduction for Claude Code sessions. To address enterprise adoption barriers, eligible customers can now opt for zero data retention, keeping their data stored locally.

The update also addresses over-sensitive safety classifiers. Anthropic reduced false positives in its biology safeguards by 85% compared to Fable 5, while Claude Code users will experience 60% fewer cyber interventions. Early feedback indicates that Fable 5.1 is highly well-rounded, showing a marked improvement in tone, a reduction in "Claude-speak," and a strong ability to simplify and delete redundant code. However, some users report that the model is highly proactive, occasionally consuming 25% to 50% more tokens at high effort levels.

Fable 5.1 demonstrates solid benchmark gains. On the Anthropic ECI, it scored 162.0, up from 159.5 for Mythos 5 and 160.7 for Opus 5. In coding and math, it achieved 67.4% on DeepSWE v1.1, a 0.57 on FrontierSWE v2, and 73.4% on CursorBench 3.2. It also improved on Humanity's Last Exam to 60.9% without tools and 65% with tools, and jumped to 52.6% on Terminal-Bench-Science 0.1 from 24.7%. On the Artificial Analysis Index, Fable 5.1 leads with a score of 57, ahead of OpenAI's GPT-6 Astra at 55, Opus 5 at 54, and Fable 5 at 53. On Vals composite, Fable 5.1 scored 68.8%, leading Astra's 66.6%.

Despite these gains, Fable 5.1 shows some regressions. It scored 0 on FrontierMath Erdos, where Astra solved 2 out of 68 problems, and hit 87.8% on FrontierMath Tier 4, slightly below Fable 5's 88%. Toolathon Verified pass@1 fell to 77.8% from 80.6%. On the Legal Agent Benchmark, it reached a 19.1% all-pass rate, up from Fable 5's 16.9%. On AutomationBench, Fable 5.1 improved to 31% from 27%, just ahead of Gemini 3.7 Flash at 30%. For developers, Fable 5.1 represents a highly capable, more affordable orchestrator for complex, long-running agentic tasks.