Anthropic cut cache costs 75% with Fable 5.1. The benchmark story is secondary.

Anthropic shipped Claude Fable 5.1 on September 1, 2026. The model is better at coding, reasoning, and identifying root causes instead of treating visible symptoms. The benchmarks moved. And the thing that actually changes how I think about building on Fable: cache reads dropped from $1.00 to $0.25 per million tokens — a 75% cut, effective immediately, no configuration required.

Cache reads are not a marginal cost in a real agentic pipeline. They're the dominant cost. Every turn in a long-running agent session re-reads the system prompt, the tool definitions, the accumulated context — all pulled from cache. Anthropic estimates typical workloads get 25% cheaper; highly agentic workloads get up to 45% cheaper. Those aren't projected benchmark numbers. Those show up in your invoice.

Base pricing is unchanged: $10 per million input tokens, $50 per million output. Fable 5.1 is available on Claude.ai, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry, for Pro, Max, Team, and Enterprise tiers.

Source spread

• Anthropic — Introducing Claude Fable 5.1 and Claude Mythos 5.1 [hype] — the company's own announcement; leads on model capability improvements; buries the cache cost change in the pricing section; worth reading the EFS detail.
• VentureBeat — 75% cost reduction for Fable cache reads [builder] — correctly identifies the cache cut as the real headline; good on EFS context.
• explainx.ai — Fable 5.1 benchmarks and pricing [builder] — granular Terminal-Bench and SWE-bench Pro breakdown; the source for 85.02% Terminal-Bench 2.1.
• 9to5Mac — Anthropic upgrades Claude with Fable 5.1 [neutral] — clean consumer-angle summary; good on the Mythos 5.1 distinction.

Pros & cons

What's real:

• The cache cost cut is structural. At $0.25/M for reads, any workload that was spending more on cache than on generation just got cheaper automatically. No API changes. Switch the model ID and the savings show up.
• SWE-bench Pro at 81.2% is meaningful. SWE-bench Pro is a harder variant of the benchmark that emerged after labs started training against the original SWE-bench Verified. An 81.2% on the harder version — when Fable 5's 95% was on the easier Verified — represents real progress, not regression. The benchmarks aren't directly comparable, but the underlying difficulty is higher and the score is still top-leaderboard.
• Mid-conversation effort adjustment in beta. If you've been managing effort by starting new sessions, this changes the workflow. One session, dial the effort up mid-run without losing context.
• Enterprise Frontier Safeguards (EFS) as an inference-layer architecture — not just a policy update. Security controls embedded at generation time. Whether it matters to you depends on your deployment, but it's the right place for them to sit.

What deserves a side-eye:

• Terminal-Bench 2.1: 85.02% for Fable 5.1 versus 90.8% for Gemini 3.8 Flash on the same benchmark, released one day later. Anthropic's frontier model trails Google's Flash-tier model on the coding-agent eval that most builders track. That's a gap worth watching.
• Knowledge cutoff is June 2026 — three months behind today. The retrieval requirement for current-events tasks doesn't change.
• The claim that Fable 5.1 is "better at identifying the underlying cause of a software problem rather than treating visible symptoms" is load-bearing marketing language. It may be true. It's also exactly the kind of claim that requires your own eval suite on real production bugs to verify.

What builders need to know

• Cache cost change is automatic. Switch the model ID, the savings appear. Check your current cache read volume against your last invoice to estimate savings — the math is 75% off that line item.
• Upgrade path is clean. Same pricing, same API surface, no breaking changes reported. Run your eval suite first and watch for instruction-following drift.
• SWE-bench Pro (81.2%) vs Verified (95%) — these are different benchmarks. Don't read the lower number as a regression. SWE-bench Pro is harder by design; the score represents genuine top-leaderboard performance on a stricter eval.
• Mid-conversation effort adjustment is beta. Useful for prototyping without session restarts; not yet production-stable.
• If Terminal-Bench is your primary criterion, run your own eval. The Gemini 3.8 Flash comparison is real and the gap isn't small.

Further reading

• Anthropic — Introducing Claude Fable 5.1 and Claude Mythos 5.1 — official launch; pricing and feature details are in the pricing section, not the lede
• Anthropic — Developing Enterprise Frontier Safeguards — EFS architecture detail
• VentureBeat — 75% cost reduction for Fable cache reads — best independent coverage of the pricing change
• explainx.ai — Fable 5.1 benchmarks — Terminal-Bench and SWE-bench Pro scoring detail
• The Decoder — Gemini 3.8 Flash Terminal-Bench context — the competitive benchmark context