Anthropic ships Claude Fable 5.1 and cuts cache reads 75%
Anthropic had a three-part Tuesday: a new frontier model, a price cut that lands squarely on agent workloads, and a retreat from the data-retention rule its enterprise customers hated.
Claude Fable 5.1 is generally available today, and the headline number is the price of context. Cache reads — what the model pays to re-read data it has already processed — now cost 25 cents per million tokens, down 75%. On token-billed usage that works out to roughly 25% cheaper overall for typical workloads and as much as 45% cheaper for context-heavy, tool-heavy agentic work, according to Anthropic's own four weeks of August usage data. Input and output pricing is unchanged at $10 and $50 per million. The capability claims are real too: 73.4% on CursorBench 3.2 at max effort, 82% on a browser-agent benchmark against 74% for Opus 5 and 57% for Fable 5, and RedlineBench jumping from 47.9 to 57.0 on contract redlining. The strategic read is the interesting part — Anthropic is not cutting headline token prices, it is cutting the cost of the loop. Agents re-read context constantly, so making the cache nearly free is a direct subsidy for exactly the workloads it wants developers to build.
Mythos 5.1 is the same model with the guards off, but only for vetted hands. It ships through two trusted-access programs aimed at cybersecurity and life-sciences work, where Anthropic's standard safeguards are restrictive enough to block legitimate research. Anthropic's own audit says Mythos 5.1 is less likely than its predecessor to reach for resources outside its test environment, less likely to reason its way around constraints, and reward-hacks at a lower rate. The science pitch is the part worth watching: the model optimized seven open-source protein and genomics models on an H100 and cut estimated GPU costs by 30 to 60% — work Anthropic says would normally take a performance-engineering team weeks and is often out of reach for academic labs. It plans to open-source those optimizations.
Apple told a federal court that OpenAI is destroying evidence, and it wants a judge to stop the hardware program. In a Monday filing, Apple says a forensic inspection turned up messages in which a former Apple employee now at OpenAI allegedly discussed "destroying the types of forensic data Apple needs," and talked about the need to "restore" and then "start using" Apple-owned devices after learning of the company's internal investigation in June. Apple is asking for a preliminary injunction blocking OpenAI from building hardware on Apple's technology, plus expedited discovery, arguing that logs, metadata and usage records are transient and "at risk of being lost, overwritten, or destroyed." The company's original complaint said more than 400 former Apple employees now work at OpenAI. Apple's case has always hinged on proving its secrets traveled; if the forensics are disappearing, the fight shifts from what was taken to whether the record can be reconstructed at all.
Worth watching: whether Anthropic's Enterprise Frontier Safeguards — free, phased in through the fall, and designed to let customers run safety scanning on their own systems instead of handing data to Anthropic — becomes the template competitors have to match. OpenAI shipped its own zero-retention option earlier this month, and enterprise privacy is now a product surface, not a policy page.
Do you think cutting the cost of cached context is a bigger lever on agent adoption than raw model intelligence? Tell us in the comments.
Sources: Anthropic · Anthropic System Card · Bloomberg via Techmeme · VentureBeat · CNBC · The Verge · TechCrunch · Apple's filing (DocumentCloud)
