Claude Fable 5.2: twice the price per token, often cheaper per task

TL;DR. On November 9, Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, its new top model under two names. It is 20% cheaper than its predecessor, but per token it still costs twice as much as Opus 5.5. We recalculated the data points Anthropic publishes in its cost charts. In four of six tests, the new model reaches the best score of Opus 5.5 for 7% to 41% of the cost per task, because it needs to think less to get there. For simple tasks, the smaller model stays cheaper.

What is new

Fable and Mythos are the same model. Fable has safeguards against misuse in cyberattacks and biology and is open to everyone: in Claude, through the API and on AWS, Google Cloud and Microsoft Azure. Mythos, without these safeguards, is only available to vetted organizations in IT security and the life sciences. According to Anthropic's announcement, the price drops from $10 and $50 to $8 and $40 per million tokens for input and output. Answers arrive up to 30% faster. A fast mode speeds things up by up to 2.5 times, at twice the price.

The biggest jump sits in a single row of the table, the Remote Labor Index. It consists of 240 real projects from freelance platforms across 23 fields of work. It counts how many results reviewers rate at least as good as a human's work. When the test was published in October 2025, the best AI managed 2.5%.

Then there are the demonstrations. 1,000 Fable agents turned the proof of Fermat's Last Theorem into a form a computer can check step by step, in 30 hours. The first attempt with a research model had taken 11 days. A team of agents built an operating system in eight days that boots on a laptop. And according to Anthropic, Claude now writes over 90% of the changes in its own core codebase, guided by humans.

Our analysis: pricier per token, cheaper per task

Fable 5.2 can think for different lengths of time on request, in five effort levels from "Low" to "Max". For six tests, Anthropic shows for each level what a task costs and how good the result is. The data points are embedded in the charts of the announcement. We extracted them and asked one question: what does it cost to reach the best score of Opus 5.5 with the new model?

For expert questions, Fable's lowest level is enough: 68.4% instead of 67.7% correct, for 18.5 cents instead of $2.18 per task. For reading charts it takes 7% of the cost, for large data-collection tasks 41%. Mythos 5.2 finds known security flaws more often at High effort than Opus 5.5 does at its highest, for a third of the cost. In coding, Fable is almost level for half the cost. Only in hardware design does it get more expensive, but also clearly better.

Two more patterns stand out. The highest level does not always pay off: in coding, Fable costs $44.70 per task at "Max" and scores lower than at "Xhigh" for $14.19. And the cheapest run comes from Opus 5.5 in all six tests. If you only need a quick, simple answer, Fable costs you extra. Anthropic itself advises starting with Opus 5.5 and only moving to Fable when Opus falls short even at a high effort level.

The catch

• These are Anthropic's numbers. The external evaluators, including the UK AI Security Institute and METR, did not review the announcement or the system card, as Anthropic itself writes. Independent rankings put Fable ahead, but by less. At Artificial Analysis it scores 65 points, Gemini 4 Carbon 60, Opus 5.5 59. In the Vals Index, Fable at 71.3 is only just ahead of Sonnet 5.5 at 70.9, which costs a quarter as much per token.
• Not ahead everywhere. In office workflows (AutomationBench), Gemini 4 Carbon leads 53.9% to 51.4%. In knowledge work, reviewers rate the results of Opus 5.5 slightly better in head-to-head comparisons.
• Blocked questions are answered by an older model. Fable blocks requests about penetration testing, exploits and most topics in research biology. The answer then comes from Opus 4.8 or Opus 5. According to Anthropic, unnecessary blocks are about 55% less frequent than with the predecessor.
• 30 days of retention. Because misuse often only shows across many requests, Anthropic monitors usage of these models over a window of time. If you have not agreed on zero data retention, your requests are stored for 30 days for this purpose.
• The model senses tests. Fable and Mythos 5.2 voice the suspicion that they are being evaluated more often than any recent Claude model. According to Anthropic, this makes it harder to predict how they will behave in real use.

Progress for anyone who lets agents work on their own systems: in more than 1,800 test scenarios with a deliberately weak sandbox, Mythos 5.2 only rarely used credentials and files placed outside it.

What this means for mid-sized companies

The price per token is misleading. A model that costs twice as much per token but finishes after a fraction of the reasoning steps pays off. That holds for hard tasks: reviewing contracts with many appendices, rebuilding legacy software, merging and cross-checking data from many sources. For classification, summaries and standard answers, Haiku, Sonnet or Opus remain the better choice. Anthropic offers an advisor mode for this: the smaller model does the work and only consults Fable on the hard cases. One date helps with planning: Fable 5.2 will not be retired before November 9, 2027.

What you should do now

1. Calculate the cost per completed task. Take 20 to 50 real cases and measure cost and result. The list price per token tells you little.
2. Start with Opus 5.5. Use Fable only where Opus falls short even at a high effort level. Anthropic gives the same advice.
3. Test the effort level. Fable runs at "High" by default. Try Low and Medium, and do not switch to "Max" blindly.
4. Bundle the hard cases. A low-cost model for the bulk of the work, Fable as an advisor for the few hard cases.
5. Clarify data protection. Ask Anthropic or your cloud provider whether the 30 days of retention apply to you, and record the answer in your data processing agreement.
6. Prepare your IT security team. If you want to use Fable for security reviews, penetration testing questions get the answer of an older model. The full scope is only available with Mythos, after vetting by Anthropic.

Fable 5.2 is not a model for everything. For hard tasks, though, it is the first where more performance can cost less. If you want to know which of your tasks belong in that group, we are happy to run the numbers on your own cases with you.