Claude Fable 5.1

The price that moved is not the one in the headline

Ten dollars in, fifty out. The same figures as Fable 5, to the cent. The only row that moved on Anthropic pricing table is the cache read: from one dollar to twenty-five cents per million tokens.

Anthropic calls this an exception itself, in a footnote under that same table. Cache reads cost 0.1x the base input price on every Claude model, and 0.025x on Fable 5.1 and Mythos 5.1 alone. In the whole price list, that is the one row that steps outside the rule.

Who feels it: anyone writing an agent. An agent resends the same system prompt and the same files on every turn, and with caching on, that column is most of the bill. Anthropic states roughly 25 percent lower cost on typical workloads and up to roughly 45 percent on agentic ones.

Now the boundary. If your workload is one long call with no repetition, there is nothing to read from cache and you pay the Fable 5 price exactly. For that shape of work, 5.1 is not cheaper. It is only better.

Where the jump is, and where it is not

Anthropic published six scored evaluations for this release. Three of them carry a Fable 5 figure as well, and those three show something a single number hides.

On agentic scientific research it goes from 24.7 to 52.6, better than double. On agentic terminal coding, 42.0 to 55.8. And on CursorBench, the closest of the three to everyday coding work, 70.5 to 73.4. Two point nine points.

All three point the same way and none of them are the same size, and that is the argument of this page. The longer and more multistep the task, the wider the gap. If your work is "write me this function", you are buying almost nothing. If your work is an agent walking a repository for forty minutes, you are buying the most this release has.

One caveat that has to stay: Anthropic measured and published all six of those numbers itself. There is no independent leaderboard behind them. We print them because the maker published them plainly and named the evaluation, not because we have confirmed them.

You do not get the 66 on the default setting

The only third-party measurement that exists for this model today is the Artificial Analysis index, where Fable 5.1 sits first of 192 models at 66. But that same page carries three numbers, not one: 66 at max effort, 65 at xhigh, and 62 at high.

The API default is high. Change the model ID and write nothing else, and the configuration you get scores 62, not 66. Opus 5 at max effort scores 63, which is above Fable 5.1 on its own default, at half the price.

We have written this before in this reference and it held again: the thing carrying a model name on a leaderboard is usually a serving configuration, not a model. To choose with, you want the number your own default settings produce.

Two more figures on that page never make a headline: output speed of 66.2 tokens per second against a median of 72.1, and a time to first token far above the median. Anthropic own table also marks this model Slower in its latency column. If your user is waiting for an answer to appear, that column matters more than any benchmark on the page.

The criticism we made of Fable 5 four months ago now has an answer

On our Fable 5 page we wrote that its knowledge stopped in January 2026 while the cheaper Opus 5 read through May 2026, and that a pricing table shows nothing of a dearer model being four months further behind. Fable 5.1 reads through June 2026. It is now ahead of Opus 5 and that criticism is settled.

A smaller change is visible in Anthropic own table: the Claude Platform on AWS column is a dash for Opus 5 and a model ID for Fable 5.1. If your company contract runs on that route, that column will matter to you more than any score.

And what did not change: the one million token window, 128k tokens of output, the input and output prices, and the absence of any official route from Iran.

Three breaking changes that cost you the first day

Anthropic migration notes call three changes breaking outright, and this is the part rarely written down, because it is not news and only costs people time.

First, forced tool use returns an error. Second, earlier models cannot read this model thinking blocks, so an architecture that swaps models mid-conversation breaks. Third, editing an earlier turn invalidates the thinking blocks.

Five further changes are additive and break nothing: per-message effort, turn-scoped system messages, readable progress updates between tool calls, that lower cache price, and content provenance. The first three are still in beta.

The newer tokenizer, and why it matters more for Persian

Another footnote under the same pricing table: models from 4.7 onward use a newer tokenizer that produces roughly 30 percent more tokens for the same text. So a dollars-per-token comparison between Fable 5.1 and, say, Sonnet 4.6 is not a comparison of like with like. The same text becomes more tokens on the newer model and the bill runs above what the price table suggests.

Fable 5 and Fable 5.1 both sit on that newer tokenizer, so comparing 5 with 5.1 is sound. Comparing either with the generation before it is not.

And here is what we do not know and will not write. Anthropic gives that roughly 30 percent for text in general and publishes no per-language figure. Persian already costs more tokens than English in most tokenizers, but we have not measured the Persian figure for this one, and until we do we will not print a number for it.

Mythos 5.1: the same model with fewer safeguards

Mythos 5.1 launched the same day. Its specifications and price match Fable 5.1 and Anthropic calls it the same capabilities, but it sells by invitation and today reaches only a set of United States organisations.

One number makes it interesting. On Terminal-Bench 4.0, Fable 5.1 scores 55.8 and Mythos 5.1 scores 60.9. The maker printed both figures side by side itself. Five point one points between a model and the same model carrying fewer safeguards is a measured cost of the safety layer, and few makers publish that number at all.