Claude Fable 5.1 nearly doubles its AutomationBench score as Anthropic cuts cache prices

Anthropic released Claude Fable 5.1 on 1 September 2026, and the number that stands out is its business-automation result: 31.4% on the AutomationBench benchmark, up from 17.1% for Claude Fable 5, which shipped roughly three months earlier. The figures and analysis below come from a dev.to write-up that compiles Anthropic's release data alongside independent pricing analysis.

What Anthropic shipped

The model arrives in two configurations. Fable 5.1 is generally available under the identifier claude-fable-5-1, reachable through the Anthropic API, Amazon Bedrock, Google Cloud Vertex AI, Microsoft Azure AI Foundry, Claude Code and Claude Enterprise. Mythos 5.1 is a restricted sibling with lighter safeguards, limited to organisations vetted through Anthropic's Cyber Verification and Life Sciences Verification programmes, currently US-only — so for most buyers, Fable 5.1 is the product.

The specification carries over largely unchanged: a 1M-token context window, 128K maximum output, and adaptive thinking that stays on. Effort tiers run from Low to Max, and 5.1 adds per-message effort control.

Benchmark gains, with caveats

According to Anthropic's published figures, Fable 5.1 leads both Fable 5 and Opus 5 on every benchmark in the release, with the largest jumps on agentic research and automation. Terminal-Bench-Science 0.1 rose from 24.7% to 52.6%; Terminal-Bench 4.0 moved from 42.0% to 55.8%; CursorBench 3.2.0 edged up from 70.5% to 73.4%; OSWorld 2.0 (strict) went from 36.1% to 41.7%; and the Elo-style GDPval-AA v2 score climbed from 1723 to 1853, ahead of Opus 5's 1824. AutomationBench puts Opus 5 at 26.9%.

Two cautions apply. The reported standard error sits at roughly 3.5 to 4.5 points, so narrow margins — CursorBench, or Humanity's Last Exam with tools at 65.0% versus 63.8% — are close to noise. Every figure is also vendor-reported. Anthropic's own comparison places Fable 5.1 well ahead of OpenAI's GPT-5.6 Sol on AutomationBench (19.6%), and launch partner Browserbase reported 82% task completion on its hardest set versus 74% for Opus 5.

Pricing: cheaper only where caching dominates

List prices are unchanged at $10 per million input tokens and $50 per million output tokens. The single cut is cache reads, down 75% to $0.25 per million. The batch API stays half price at $5/$25, and cache writes remain $12.50 per million for the 5-minute window and $20 for the 1-hour window. Anthropic estimates typical workloads get about 25% cheaper and complex agentic tasks up to roughly 45% cheaper.

Independent analysis by Stork.AI complicates that: Fable 5.1 reportedly emits about 1.7x more output tokens per task, and at Max effort some analyses put cost per task around 20% higher than Fable 5 despite the cache discount. Because output tokens bill at $50 per million, the dev.to write-up concludes that 5.1 wins on economics only when cached context dominates spend — a stable codebase or knowledge base hit thousands of times, or long agent sessions that keep revisiting the same material. For varied one-off prompts, Opus 5 at $5/$25 or Sonnet 5 at $2/$10 remains better value. Reviewers at The Decoder and VentureBeat reportedly reached the same conclusion: choose Fable 5.1 when finishing the task matters more than minimising token cost. Anthropic itself noted Fable 5 captured only about 11% of model spending across 70,000 companies, with cheaper rivals taking share.

Longer unattended runs, fewer shortcuts

The automation-relevant claims extend beyond scores. Launch partners reported MongoDB building a working prototype over roughly three days of unattended work, Ramp running a 38-hour unattended machine-learning training job with its own evaluation loop, and Millennium tracing a rare crash to a third-party vendor library after the bug had resisted explanation for four to five years. Anthropic also says the model mapped dependencies across 8 services and 3 codebases in a single multi-repo task.

Anthropic further claims 5.1 avoids reward-hacking-style shortcuts such as planting test answers in code or signalling success falsely — behaviours earlier models sometimes used to look finished. That matters as much as any score in production: an agent that quietly fakes a passing test ships its failure, while one that fails loudly can be retried.

Why it matters

Nearly doubling an automation benchmark within one three-month release cycle is a fast rate of progress on precisely the capability that gates real deployments — finishing multi-step work without a person intervening. The pricing story is subtler than the 75% cache cut implies, since higher output-token usage can erase the saving, which shifts the useful comparison from price per token to cost per completed task. Buyers should also weigh the sourcing: the benchmark numbers are Anthropic-reported and relayed here through a single secondary write-up, and the independent token-usage data points the other way on cost.