Fable 5.1 holds its price and cuts your bill — the cache-read lever explained

Read the pricing page for Claude Fable 5.1 and you would conclude nothing had changed: $10 per million input tokens, $50 per million output, exactly as before. Read a month of your own agentic billing and you would find something quite different, because Anthropic cut the line that dominates that bill by three quarters.

In long-running agentic work, the same context gets re-sent constantly. A coding agent working through a repository carries the same system prompt, the same file tree, the same accumulated history into every turn. Providers cache that repeated prefix, and the cache-read rate — not the headline input rate — is what most of those tokens are billed at. Cut it 75% and the economics of long-horizon work change without a single headline number moving.

Anthropic quantifies the effect, and is careful to label it an estimate: around 25% less than Fable 5 for typical workloads, and often much larger for highly agentic work — up to approximately 45%. Those figures come from the company's own four-week usage measurement in August 2026, so they describe a particular workload mix rather than a guarantee. The direction is not in doubt; the magnitude for your application is an empirical question.

The release pairs the cache cut with a rebalanced effort ladder. At Low or Medium effort, Anthropic says Fable 5.1 achieves results similar to or better than Fable 5 at much lower cost; at higher effort it reaches capability its predecessor could not. Defaults reflect that split — High in Claude Code, Medium in Claude Cowork and on Claude.ai.

That positioning matters competitively. Two days later OpenAI launched GPT-6 Astra at the same headline $10/$50 — but, as several analysts noted, without matching Anthropic's cache-discount lever, which leaves Astra at a relative disadvantage for high-cache-reuse applications. When two flagships carry identical sticker prices, the cache rate becomes the real differentiator for anyone running agents.

The two names are the same model. What differs is what the model is permitted to do, and for whom.

Mythos runs through two trusted access programmes aimed at vetted cybersecurity and life-sciences organisations that need capabilities production safeguards normally constrain. Under Fable, a request that trips classifiers around cybersecurity, biology and chemistry, or model distillation is handled by a less capable Claude Opus instead. It is a tiered-access model of safety rather than a single global refusal boundary — and as an approach it is quietly becoming the industry norm: OpenAI's Daybreak programme does something structurally similar for cyber capability.

Anthropic calls these the world's most advanced models for coding and knowledge work, and says Fable 5.1 outperforms Fable 5, Opus 5 and GPT-5.6 Sol across multiple benchmarks. The most striking specific claim is not a coding number at all: on three targets, its protein binding affinities were ten times higher than the best designs submitted to Adaptyv Bio's protein design competitions. That is a concrete, externally-anchored result of a kind benchmark tables rarely produce.

Two caveats belong alongside it. Anthropic acknowledges that production safeguards can affect scores — meaning the Fable you use may not perform identically to the Mythos that was measured — and that its August 2026 OSWorld task release is not directly comparable with some previously published results. Those are the right disclosures to make, and they are also a reminder that cross-lab benchmark comparisons in this cycle are shakier than the tables imply.

What to watch: whether OpenAI answers the cache-read lever on Astra, whether the effort-level defaults get tuned again once usage data accumulates, and whether Mythos-class trusted access broadens beyond cyber and life sciences. The pricing move is the one to copy, though — cutting the cache rate rather than the sticker price is how you lower an agentic bill without repricing your flagship.