Claude Fable, the model above Opus

Anthropic explains the name itself: “Fable is from the Latin fabula, ‘that which is told,’ akin to the Greek mythos.” Mythos is the model Anthropic held back in April because of what it can do in cyber security. Fable is the same model with safeguards, the version it is willing to tell everyone about.

Fable 5 arrived on 9 June, Fable 5.1 on 1 September. Here is where it sits, what it is for, what it costs, how fast it is and where it falls short.

Where it sits

Anthropic now has four lines of Claude: Haiku, Sonnet and Opus, and above them the Mythos class, sold in two forms. “Claude Fable 5.1 and Claude Mythos 5.1 are the same model, but with different levels of safeguards.” Fable 5.1 is on general sale. Mythos 5.1 is open only to organisations verified through Anthropic's programmes for cyber security and life sciences, so far in the US.

Anthropic's documentation describes Fable 5.1 as the model “for demanding reasoning and long-horizon agentic work”, and lists its speed as slower than the rest of the lineup. It reads text and images, holds a million tokens of context, writes up to 128'000 tokens in one answer and thinks before every answer, by default at high effort. It runs on Anthropic's platform, Amazon Bedrock, Google Cloud and Microsoft Foundry.

It also comes with extra rules. Amazon notes that Anthropic has designated Fable 5.1 a Covered Model, which carries additional data retention, safety review and access policies wherever it is offered. Until a scheme that keeps the data in a cloud environment the customer controls rolls out from this autumn, eligible customers can use it with zero data retention.

What it is for

Fable is built for long work that runs without a person at every step. Amazon describes it handling features across an entire codebase, code review and performance work over sessions that last hours, with fewer confident wrong answers than Fable 5. One line from that announcement matters more than any benchmark: “If it gets stuck, it says so instead of reporting success”.

Anthropic's own figures show the jump from Fable 5. On Terminal-Bench 4.0, work in a command line, it rose from 42.0 to 55.8 percent; on AutomationBench from 17.1 to 31.4 percent; on scientific tasks in the terminal from 24.7 to 52.6 percent. At the launch of Fable 5, Anthropic reported that Stripe completed a migration of 50 million lines of Ruby in a day, work it said would have taken a team more than two months.

It is also being used for research. Anthropic says Fable 5.1 trained a neural network to build a new, sharper elevation map of a third of Venus from radar data collected 30 years ago, and released the map under a Creative Commons licence.

Price and speed

Fable 5.1 costs 10 dollars per million input tokens and 50 dollars per million output tokens, the same as Fable 5 and two and a half times the price of Opus 5.5. What changed is cached input, which fell by 75 percent to 0.25 dollars per million tokens. Anthropic estimates savings of about 25 percent on typical work and up to about 45 percent on highly agentic work.

It is slow and it is thorough. Artificial Analysis ranks it fifth of 216 models on its intelligence index, but measured about 71 tokens a second at maximum effort, below the median for its class, and a long wait before the first token while it thinks. It also used 190 million output tokens to run the index, against a median of 85 million. Zvi Mowshowitz, who follows these releases closely, collected reports of heavy token use and suggests lowering the effort setting to manage it.

Where it falls short

The safeguards are part of the product. Classifiers watch for requests about cyber attacks, biology and chemistry, and attempts to copy the model, and hand them to an older model. With 5.1, Anthropic says the cyber safeguards produce 60 percent fewer false alarms, and Fable can now help find software vulnerabilities, but not write exploits for them. Some users still report being moved to the older model in the middle of a session.

Availability carries a risk of its own. On 12 June the US Department of Commerce required an export licence for any foreign person to use Fable 5, and Anthropic switched the model off for every customer until 1 July. For a company that had built its work on Fable, that meant nearly three weeks without its best model.

And its lead has narrowed fast. Three weeks after Fable 5.1, Opus 5.5 beat it on most of Anthropic's own tests at less than half the price, although Anthropic says the gap in its own use is narrower than the scores suggest.

Our verdict

Fable 5.1 is an excellent model for a narrow set of jobs: a migration across a large codebase, a long research task, an audit that must not miss anything, work that runs for hours and where a wrong answer costs more than the wait. For daily work, Anthropic's own advice is the right one: start with Opus 5.5, and move up only where your own tests show that it falls short.

Whichever model wins, the knowledge should stay with the company. Twin gives Claude, ChatGPT, Codex or a private model the company's knowledge, rules and skills, so moving between Opus and Fable is a setting, not a project. For work that must not leave the building at all, Eigna, built by our partner Eigna Labs, is a private model trained on a company's own work and running in its own cloud or on its own servers. It is designed to run offline, so a supplier's decision cannot simply switch it off.