← The Polylog AI Intelligence Brief
Morning Edition · Saturday, September 12, 2026Published at 2:23 AM EDT · New York
Anthropic Cuts Claude Fable 5.1 Cache Reads by 75 Percent
Cache reads now cost $0.25 per million tokens, which Anthropic says reduces typical workloads by about a quarter and heavily agentic workloads by up to roughly 45 percent.
Anthropic has released Claude Fable 5.1 and Claude Mythos 5.1, and the change that will show up fastest in production bills is not a benchmark score but a pricing change. Base rates hold at $10 per million input tokens and $50 per million output tokens, but cache reads drop to $0.25 per million, a 75 percent reduction from Fable 5. VentureBeat reports that the company estimates this cuts typical workload cost by about 25 percent and highly agentic workload cost by up to roughly 45 percent.
That difference in treatment is the point. An agent loop re-reads a large, stable block of context on every turn, so cached input dominates its token bill in a way it does not for a one-shot chat exchange. By repricing the cache read rather than the base rate, Anthropic is targeting a subsidy specifically at long-running agents, which moves the economics of tool-heavy systems more than a headline price cut of the same nominal size would.
The two models are the same underlying system at the same price, separated by safeguards and access. Fable 5.1 is generally available with added protections around cybersecurity, biology, and chemistry. Mythos 5.1 runs without those restrictions and is limited to a small set of vetted organizations, currently based in the United States, in Anthropic's trusted access programs. The cost of those safeguards in raw capability is visible in Anthropic's own numbers: on Terminal-Bench 4.0, Mythos 5.1 scores 60.9 percent against 55.8 percent for Fable 5.1.
Anthropic says the release leads Claude Opus 5 on every benchmark it published, reversing results in which Opus 5 previously scored higher. Those are vendor-run evaluations on a vendor-selected set of benchmarks, and no independent reproduction of the Terminal-Bench figures has appeared yet.
What this means
Cache-read pricing is becoming the main point of competition for agent workloads, because the marginal token in an agent run is overwhelmingly a re-read of context rather than a fresh prompt. Vendors that price the cache aggressively capture the long-horizon coding and research workloads where per-task token counts run into the millions, and vendors that hold cache rates close to input rates lose those workloads regardless of where they rank on a capability leaderboard. The five-point Terminal-Bench gap between the safeguarded and unsafeguarded versions also puts a specific number on what safety filtering costs in coding performance, and that is information both regulators and buyers will use.
Part of a tracked trend
Frontier Model Price War
Frontier API vendors increasingly compete on strategic price-cutting rather than pure capability, repeatedly launching or repricing models below prevailing rates to grab share and compressing industry inference margins; expect recurring below-rival pricing moves.
Townsquare
Start a discussion in Townsquare.
More from this edition
- Clay Institute Declines to Call Navier-Stokes Solved After OpenAI's AI Proof Claim
- Anthropic Says DeepSeek and Moonshot Routed Customer Prompts to Claude Through Fake Accounts
- Perplexity and Cognition Put GPT-6 Astra on Production Systems
- Anthropic Hands METR an Eight-Week Review After Claude Models Reached Live Systems
- Meta Claims Coding Leadership for Muse Spark 1.3, With Its Best Numbers From a Restricted Variant
- Sber Open-Sources a 432-Billion-Parameter Reasoning Model Under MIT License
- Sakana AI Splits Its Router Into a Cheap Tier and a Capability Tier
Comments
0No comments yet.