Claude Fable 5.1

The parameter that changed is not the one in the headline

Ten dollars in, fifty out. The same values as Fable 5, without a cent of difference. The only field that shifted on Anthropic's pricing table is the cache read rate: from one dollar to twenty-five cents per million tokens.

Anthropic itself designates this an exception, logged in a footnote under that same table: the cache-read multiplier across every Claude model is 0.1x the input rate, and only for Fable 5.1 and Mythos 5.1 is that multiplier set to 0.025x. Across the entire pricing list, this is the one row that sits outside the standard rule.

Which workload feels this: an agentic one. An agent resends the same system prompt and the same file set on every execution turn, and with caching enabled, that field accounts for the bulk of the bill. Anthropic reports roughly 25 percent lower cost on typical workloads and up to roughly 45 percent on agentic ones.

And now the boundary of that reduction: if your workload is one long call with no repetition, there is no data to read from cache, and you pay the Fable 5 rate exactly. For that workload shape, 5.1 is not cheaper — it is purely a performance improvement.

Where the jump occurs, and where it does not

Anthropic published six scored evaluations for this release. Three of them carry a corresponding Fable 5 value, and those three expose something a single figure conceals.

On agentic scientific research the value climbs from 24.7 to 52.6, more than double. On agentic terminal coding, from 42.0 to 55.8. And on CursorBench, the closest of the three to everyday coding workloads, from 70.5 to 73.4 — a gap of two point nine points.

All three values move in the same direction, but their magnitude is not remotely equivalent, and that is the central finding here. The longer and more multistep the task, the wider the performance gap. If your workload is "write this function", you are purchasing almost no improvement. If your workload is an agent traversing a repository for forty minutes, you are capturing the largest gain this release delivers.

One constraint that has to stay on record: Anthropic measured and published all six of these values itself. No independent leaderboard stands behind them. They are printed in this reference because the maker published them plainly, with the evaluation named, not because they have been independently verified.

The value of 66 is not reached on the default configuration

The only independent third-party measurement available for this model today is the Artificial Analysis index, where Fable 5.1 holds rank one of 192 models at 66. But that same page reports three values, not one: 66 at the max configuration, 65 at xhigh, and 62 at high.

The API default is provisioned at high. That means if you swap only the model ID and set no other parameter, the configuration you receive scores 62, not 66. Opus 5 at the max configuration logs 63, which sits above Fable 5.1's own default, at half the rate.

This has been logged in this reference before, and it held again: a value carrying a model's name on a leaderboard is usually a serving configuration, not the model itself. For an operational decision, the value produced by your own default configuration is what carries weight, not the peak figure.

Two further values on that page never make a headline: output throughput of 66.2 tokens per second against a median of 72.1, and a time-to-first-token figure well above the median. Anthropic's own reference table also logs this model's latency field as "Slower". If your system's end user is waiting for an answer to appear, that field is more operationally critical than any benchmark value on the page.

A limitation logged against Fable 5 four months ago is now resolved

In the Fable 5 analysis, it was logged that its knowledge cutoff was January 2026 while the cheaper Opus 5 shipped with a May 2026 cutoff, and that a pricing table shows no trace of a dearer model trailing by four months. Fable 5.1's knowledge cutoff is set at June 2026. It now sits ahead of Opus 5, and that limitation is resolved.

A smaller change is also visible in Anthropic's own reference table: the Claude Platform on AWS field is a dash for Opus 5 and carries an active ID for Fable 5.1. If your organization's infrastructure holds a contract on that route, this field carries more operational weight than any score.

And the parameters that did not change: the one-million-token context window, the 128k-token output ceiling, the input and output rates, and the absence of an official deployment route from Iran.

Three breaking changes that stall day-one deployment

Anthropic's migration documentation explicitly designates three changes as breaking, and this is the part that gets documented less often, because it is not newsworthy and only consumes engineering time.

First, forced tool invocation returns an error. Second, earlier models cannot parse this model's thinking blocks, so any architecture that swaps models mid-conversation fails. Third, editing an earlier conversation turn invalidates the thinking blocks.

Five further changes are additive and break nothing: per-message effort configuration, system messages scoped to a single turn, readable progress reporting between tool calls, the lower cache rate itself, and content-provenance tagging. The first three remain in beta.

The new tokenizer, and why it carries more weight for Persian

Another footnote under the same pricing table: models from version 4.7 onward run on a new tokenizer that produces roughly 30 percent more tokens for identical text. That means a dollars-per-token comparison between Fable 5.1 and, say, Sonnet 4.6 is not a comparison of equivalent units; the same text yields more tokens on the newer tokenizer, and the actual bill runs above what the pricing table implies.

Fable 5 and Fable 5.1 are both deployed on this same newer tokenizer, so comparing 5 against 5.1 is valid. Comparing either against the prior generation is not.

And here is what has not been measured in this reference, so no figure is logged for it: Anthropic reports that roughly 30 percent for text in general and publishes no per-language figure. Persian already carries a higher token cost than English on most tokenizers, but the specific Persian figure for this tokenizer has not been independently measured, and no number will be printed until it is.

Mythos 5.1: the same model, a lighter safety layer

Mythos 5.1 also went into production the same day. Its technical specification and pricing rate match Fable 5.1 exactly, and Anthropic describes it as carrying "the same capabilities", but its distribution is invitation-only, reaching today only a set of United States organizations.

One value makes this comparison notable: on Terminal-Bench 4.0, Fable 5.1 logs 55.8 and Mythos 5.1 logs 60.9. The maker published both values side by side itself. The five-point-one-point gap between a model and that same model with a lighter safety layer is a measured cost of the safety layer itself, a figure few makers publish at all.