Claude Fable 5.1: Pricing, Features & Review (2026)

Update (9 Nov 2026): Claude Fable 5.2 and Mythos 5.2 are out - prices, access and benchmarks are in my Claude Fable 5.2 guide.

👁 118 reads

Claude Fable 5.1 is Anthropic’s most expensive and powerful model, released on September 1, 2026, alongside its twin Mythos 5.1. In brief: it costs $10 per million input tokens and $50 per million output tokens, reads up to one million tokens of context at once, and is aimed at tasks where only quality matters - difficult code and lengthy research work. For writing and editing in non-English languages, the twice-cheaper Opus 5 is more than sufficient, as we confirmed in our own test described below.

This article covers everything you need to know before using this model: current pricing, real differences from its predecessor, the rationale behind the Fable and Mythos split, availability, and results from our launch-day testing. We update it as new information emerges.

What is Claude Fable 5.1?

It’s a refreshed version of the model that debuted on June 9, 2026, which Anthropic describes as its best tool for programming and knowledge work. The Fable family sits at the top of this company’s offering, above the cheaper Opus and Sonnet lines, and is designed for the most difficult tasks rather than everyday query handling.

Version 5.1 is not a ground-up rebuild. Changes cover three areas: quality in code and research, safety filter behavior, and two new experimental features. The fundamentals - price and context window - remain untouched.

How much does Fable 5.1 cost?

API pricing looks like this:

• $10 per million input tokens,
• $50 per million output tokens,
• context window of 1,000,000 tokens, maximum response 128,000 tokens.

This is identical pricing to Fable 5, so the launch doesn’t increase costs for anyone already working with this family. For comparison, in the same Anthropic offering, Claude Opus 5 costs $5 and $25, while Claude Sonnet 5 costs $2 and $10. Fable is therefore twice as expensive as Opus and five times more expensive than Sonnet.

Anthropic states that the models are 25 percent cheaper to run for most customers, and cache reads are up to 45 percent cheaper for complex tasks. These are manufacturer claims and should be treated as such until independently verified.

Fable 5.1 compared to the rest of the lineup

Model	Input (USD per million)	Output (USD per million)	Context window
Claude Fable 5.1	10	50	1 million
Claude Fable 5	10	50	1 million
Claude Opus 5	5	25	1 million
Claude Sonnet 5	2	10	1 million
Claude Haiku 4.5	1	5	200k

What changed compared to Fable 5?

Most practical are two features in experimental mode. The first is effort adjustment during conversation: you can ask the model to work harder on one specific step without starting a new session. For people conducting long working conversations, this is a real time-saver, since previously the only way to change modes was usually to start over.

The second is content provenance marking - tagging responses so their source can be verified later. It sounds like a technical detail, but it’s a feature publishers and legal departments have been requesting for months.

The third change concerns safety. Anthropic reports that the filter generates 60 percent fewer false alarms, and the boundary now runs between searching for vulnerabilities, which the model agrees to, and building a working attack tool, which it refuses.

Fable 5.1 or Opus 5? We tested it ourselves

On launch day, we ran Fable 5.1 through the same set of six writing tasks we’ve been using to measure models since August: declension of foreign proper names, collective numerals, capitalization, participial clauses, commas, and rewriting official language into plain text. We tested in Polish, a heavily inflected language, as a proxy for non-English writing quality. We placed Fable 5 and Opus 5 alongside it.

Results from one run (we tested it in Polish):

• Fable 5.1: five tasks out of six. On the numerals task, the model didn’t respond at all, hanging until the connection was interrupted after three minutes. Average 6.5 seconds per response, entire set cost 7 cents.
• Fable 5: six out of six, average 7.2 seconds, cost 9.6 cents.
• Opus 5: six out of six, average 8.2 seconds, cost just under 6 cents.

It’s worth knowing that freezing on a short language task isn’t new in this family. Fable 5 behaved identically in the August measurement, just on a different task. An interesting note in the other direction: in the capitalization task, both Fable models corrected the text according to Polish norms, while Opus 5 left certain phrases capitalized incorrectly, committing the most common error in Polish official documents.

Practical conclusion: for writing and editing in non-English languages, the twofold price difference isn’t reflected in results. Fable justifies itself elsewhere - in code and long research tasks, which this set deliberately doesn’t touch. We describe the full results and methodology in a separate article about the launch-day test.

How does Fable 5.1 differ from Mythos 5.1?

Not in power, but in boundaries. According to the manufacturer, the same machine works underneath, and what differs is the list of things the model won’t do. Fable is available to anyone who pays and refuses topics related to computer attacks and dangerous biology. Mythos goes only to verified recipients in cybersecurity and life sciences who need these same topics for their work.

This split has a concrete background. In June 2026, after researchers bypassed Fable 5’s safeguards, the U.S. Department of Commerce ordered access to the models cut off for people outside the United States. Both models were disabled worldwide for 19 days and returned on June 30. Releasing the same machine in two filter variants is a direct response to that lesson. We break this down further in an article about one model released with two sets of rules.

Does Fable 5.1 work outside the US?

Yes. The model is available in paid Claude plans and through the API, including for international users, and works the same way through resellers offering access to multiple models at once. However, the June incident showed that availability of this particular family can depend on administrative decisions in the U.S., not just whether you’ve paid for a subscription.

The practical takeaway for businesses is simple: if you’re basing a process on this model, keep the ability to switch to another with a single setting in your code, and calculate how much the same work costs on a model from the second tier.

What Fable 5.1 does best and where it falls short

Best at: difficult code, multi-hour agent work, analysis of extensive documents using the million-token window, research tasks requiring juggling multiple threads at once. These are applications where the higher per-token rate is often the cheapest line item in the entire bill, because the alternative is a specialist’s time.

Worse at: everything that’s short and repetitive. For simple language tasks, you pay twice as much for a result that a cheaper model delivers equally well, and sometimes better. In our measurement, Fable 5.1 even used 675 output tokens for a response identical character-for-character to one that Fable 5 returned in 359 tokens. For one sentence that’s pennies; for thousands of daily queries, it’s a budget line item.

How to get access

There are three routes. First is the Claude app in a paid plan, where you select the model from a list. Second is Anthropic’s API, billed according to the rates given above. Third is aggregators offering multiple models in one interface, convenient when you want to compare models or have a backup ready in case of outages.

Mythos 5.1 requires a separate path: application and verification in the trusted recipient program, intended for those working in cybersecurity and life sciences.

Is it worth it?

If you write code, conduct long agent tasks, or work with documents that no other model can fit, then yes, and then 5.1 is the natural choice instead of the older version, since it costs the same. If your work is writing, editing, and research in non-English languages, you’re buying a twice-more-expensive ticket for a journey that a cheaper model performs equally well.

One honest caveat to close: independent benchmarks haven’t yet measured version 5.1. On launch day, the Artificial Analysis index, which we use in our model ranking, had only Fable 5 in the table with a score of 62 points. Until independent measurements appear, all we have are manufacturer claims and our own trials. We update this article as soon as the numbers emerge.

Fable 5.1 versus competing models

Comparisons between model families age fastest of everything we write, because each launch resets the stakes. So instead of a table that will be wrong in a month, we stick to two things that don’t change.

First is price positioning. Fable 5.1 is one of the most expensive publicly available models on the market and was designed that way. If your task can be done with a mid-tier model, do it with that, because the difference in cost can be several-fold, and in many applications you won’t see a difference in results.

Second is how to check. Instead of trusting manufacturer charts, it’s worth taking five of your own typical tasks and running them through two or three models, measuring time, cost, and whether the answer is usable without corrections. Such a test takes fifteen minutes and tells you more about your work than any general benchmark. That’s exactly how we built the non-English language test in which the most expensive model failed the numerals task.

We collect current model positions in the independent intelligence index, along with rates and speed, in our AI model ranking, which updates automatically.

Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →

Reader questions

How much does Claude Fable 5.1 cost?

API rates are $10 per million input tokens and $50 per million output tokens. That’s exactly the same as the older Fable 5, and twice as much as Claude Opus 5 ($5 and $25). In the Claude app, the model is included in paid plans, and in the Max plan it works without surcharges up to the plan limit.

How does Fable 5.1 differ from Fable 5?

Price, the million-token context window, and the 128,000-token response limit remain unchanged. Anthropic claims better results in coding and research work, fewer false alarms in the safety filter (by 60 percent), and two features in experimental mode: effort adjustment during conversation and content provenance marking.

Is Fable 5.1 better than Opus 5?

In programming tasks and long research work, Anthropic positions Fable higher. In our own non-English language test on launch day, the difference didn’t justify the twice-higher price: Fable 5.1 solved five tasks out of six, didn’t respond at all to one, while Opus 5 completed all six and cost the least.

How does Fable 5.1 differ from Mythos 5.1?

It’s the same model underneath, differing in what it allows. Fable is publicly available and refuses topics related to computer attacks and dangerous biology. Mythos goes only to verified recipients in cybersecurity and life sciences who need these topics professionally.

Does Fable 5.1 work internationally?

Yes, the model is available in paid Claude plans and through the API, including for international users. However, it’s worth remembering that in June 2026 the previous version was disabled outside the United States for 19 days following an administrative decision, so availability of this model family has sometimes been geography-dependent.

What is Fable 5.1’s context window?

One million input tokens, roughly enough to fit several thick books at once. Maximum length of a single response is 128,000 tokens.

We also collect technical specifications, version history, and comparisons in the Fable 5.1 card in our model atlas. We describe the older version separately in an article about Fable 5, and the entire history of this family in a topic hub.

Kinetyra