Claude Fable 5.1 cuts cache reads to $0.25 per million

Anthropic shipped Claude Fable 5.1 and Mythos 5.1 on September 1, and the headline is not the model. It is one number in the pricing table: Fable 5.1 cache reads cost $0.25 per million tokens, down 75% from $1. Base input stayed at $10/M and output at $50/M. Nothing about the sticker price changed. Everything about the prompt caching math did.

What actually happened

Anthropic's pricing docs make the mechanism explicit. Every other Claude model prices a cache hit at 0.1x base input. Fable 5.1 and Mythos 5.1 price it at 0.025x — a special-cased multiplier that exists nowhere else in the lineup. Five-minute cache writes are unchanged at $12.50/M (1.25x), one-hour writes at $20/M (2x).

Anthropic's own framing in the launch post is that Fable 5.1 runs about 25% cheaper than Fable 5 on typical workloads, and up to ~45% cheaper on highly agentic work. That gap is the whole story. The 45% case is not a better model — it is a workload where the same system prompt, tool schemas, repo context and prior turns get resent on every hop, and now cost a quarter of what they did. Fable 5.1 is available today on the Claude API as claude-fable-5-1, plus AWS, Google Cloud and Azure.

Worth keeping in frame: Claude Opus 5 is still $5/M input, $25/M output, $0.50/M cache reads. On absolute dollars Opus is cheaper. The crossover depends entirely on your cache hit rate, not on anyone's blog post.

Why cache pricing matters for your business

Most teams we onboard cannot tell us what fraction of their input tokens are cache hits. They can tell us their monthly bill and their model name, and that is it. That is a problem now, because Anthropic just made cache hit rate the single largest lever on an agent's unit cost — and made it model-specific, so the answer changes when you switch.

The Messages API already hands you the numbers. Every response carries cache_read_input_tokens and cache_creation_input_tokens in usage. Log both, per run, next to the model ID and the task outcome. Then compute hit rate as cache reads over total input. Under roughly 50%, the discount is noise and you should be tuning breakpoints, not celebrating. Above 80%, the 0.025x multiplier is doing real work and Fable 5.1 may genuinely undercut a nominally cheaper model.

The second thing to fix is cache-hostile prompt construction. Timestamps, request IDs, shuffled tool lists and per-user preamble at the top of a prompt invalidate everything after them. Put the stable material first — system instructions, tool schemas, documentation — set your breakpoint after it, and keep the volatile bits at the tail. A 5-minute write costs 1.25x; it pays for itself after a single read. If your agent loops for minutes at a time, the one-hour cache at 2x is usually the cheaper write.

Sources: Anthropic, Claude Platform pricing docs.