Claude Fable 5.1 for developers: does the 75% cache-read cut actually change your agent loop?

What it is

Claude Fable 5.1 is Anthropic's point release on top of Fable 5, shipped September 1. Claude Fable 5.1 and Claude Mythos 5.1 are the same model — Mythos 5.1 is the identical model with lighter safeguards, restricted to vetted organizations through the Cyber Verification Program and the Life Sciences Verification Program, currently US-only. For most developers, Fable 5.1 is the one you'll actually call.

The headline for our persona isn't a benchmark bump — it's the cache economics. Anthropic has cut a Fable 5.1 cache hit to just $0.25 on input, down from $1.00 for Fable 5. That's also just 2.5% of Fable's normal input-token price of $10, rather than the 10% multiplier used by most other Claude models.

The next-work-session test

You're iterating on a coding agent. Same system prompt, same tool schema, same repo tree stuffed into context — you run the harness dozens of times a day tweaking the planner. Under the old 10% cache-read multiplier, every eval sweep was a line item. Under the new 2.5%, the cached prefix is close to a rounding error, and you stop rationing runs.

Concretely: if your cached prefix is 200k tokens, a cache hit now bills at $0.05 for that prefix instead of $0.20. Across a 100-run eval sweep, that's the difference between "ask for approval" and "just run it." Anthropic claims cache reads fall 75% to $0.25 per million tokens, cutting typical workload costs by about 25% and agentic ones by up to 45%. Their number, their workloads — but the direction is right.

Pricing

Verified via Anthropic's pricing docs. On Claude Fable 5.1 and Claude Mythos 5.1, a cache hit costs 2.5% of the standard input price ($0.25 USD per million tokens). These multipliers stack with other pricing modifiers, including the Batch API discount and data residency. Base input remains $10/M and output $50/M per the DataCamp writeup of the same tier. Cache writes are unchanged; the discount is read-side only.

What we'd actually use it for

• Eval harnesses with large, stable prefixes. System prompt + tool defs + few-shot corpus cached once, then hammered by parametric test cases.
• Multi-turn agent loops where each tool-call round re-sends the growing transcript. The 512-token cache minimum is low enough that most real agent contexts qualify — it keeps the 1M token context window by default, 128k max output tokens, the 512-token prompt caching minimum, and mid-conversation system message support.
• Repo-grounded coding agents. Load the tree once, cache it, iterate.
• Drop-in from Fable 5. Claude Fable 5.1 uses the same Messages API and tool use patterns as Claude Opus 5. Migration is a model-ID swap.

Not a use case: swapping in Fable 5.1 for a one-shot chat call. No cache, no discount.

Limits

• Cache reads only. The 75% cut doesn't touch cache writes, base input, or output. If your agent burns tokens on long generations rather than long prompts, this changes little.
• Cache TTL still applies. If your eval sweep is spread across hours with idle gaps, you're paying to re-warm the cache. Batch your runs.
• max_tokens accounting. Max_tokens is a hard limit on thinking plus response text together. At high effort you need room for both, which is why the agent loop below sets 16,000 rather than something tidier. Extended thinking eats your output budget — plan for it in the SDK call.
• Mythos is gated. If you were hoping to test the less-restricted variant against a security eval, Fable 5.1 is available to all Claude Pro and Max users, while Mythos 5.1 is part of a trusted access program designed for scientists and cybersecurity researchers — regular users can't access it through public subscriptions.
• Vendor benchmarks are vendor benchmarks. The 45% agentic-cost-reduction figure is Anthropic's number on Anthropic's workload mix. Measure your own.

Try it if

• You run repeated evals against a stable, large prompt prefix.
• Your agent loop already uses prompt caching and you have the cache-hit metrics to prove it.
• You're on Fable 5 today — the migration guide is short and the API surface is identical.
• You want the 1M-token context for repo-grounded work without paying Opus-class rates on every turn.

Skip it if

• Your workload is short prompts and long outputs — cache reads aren't your bottleneck.
• You're happy on a smaller model (Haiku-class) and the eval quality is already fine. Fable 5.1 base pricing hasn't moved.
• You need Mythos 5.1 access but aren't in a verified cyber or life-sciences program — you won't get it.
• You haven't instrumented cache-hit rate. Turn that on first, then decide.