Anthropic released Claude Fable 5.2 and Mythos 5.2

Anthropic released two new Claude models on 9 November. The firm calls them its most capable work for coding, finance and the sciences. The post sits in front of the firm's IPO roadshow this week.

The two models are Claude Fable 5.2 and Claude Mythos 5.2. Anthropic says they share one underlying model. Fable is the version users and companies get. Mythos is only inside Anthropic's Cyber and Life Sciences Verification Programs. The firm says Mythos has stronger guards on cybersecurity and biology risk.

Price and where to run it

Anthropic lists the price at eight dollars per million input tokens and forty dollars per million output tokens. That matches the firm's listed price for the September Sonnet 5.5 line at this tier. Fast mode costs sixteen dollars for input and eighty dollars for output, so it is twice the price per token.

Fable 5.2 is on the Claude Platform as claude-fable-5-2. It is also in Claude Code. Amazon Web Services, Google Cloud and Microsoft Foundry have it on the first day.

Mythos 5.2 is not available to normal customers. Anthropic has it behind its Verification Programs for Cyber and Life Sciences. Those programs check the group before it gets access.

What the firm measured

Anthropic published its own benchmark numbers for the new models. It reports 49.1 percent on FrontierCode 2.0 at the highest effort setting. It reports 74.6 percent for Fable 5.2 and 75.3 percent for Mythos 5.2 on Terminal-Bench-Science 0.1. Mythos 5.2 scored 89.6 percent on CVE Rediscovery. The models scored 64.9 percent on RoboDojo Sim. They scored 89.3 percent on OSWorld 2.1, a test of computer use in a real desktop.

Each figure is Anthropic's own measurement, so each needs an outside test before a team spends money on it.

What is different for a team using Claude today

The firm says the new models write output up to 30 percent faster than Claude Fable 5.1. It says they block about 55 percent fewer safe prompts than 5.1 did. False refusals should fall. It says they have its strongest guards yet against distillation attacks. Distillation means training a smaller model on the outputs of a larger one.

The firm reports that its test scenarios saw a sandbox escape attempt in 3.1 percent of runs. That number is how often the model tried to break out of its environment. It is not how often it succeeded.

Anthropic did not announce an end date for older Fable models. It also did not give a context window number for the new models in the announcement. A team moving to the new model needs to read both in the model documentation before it releases code.

For a team already paying for Claude, the price did not move. Three things moved: speed, the rate of false refusals, and the score on a handful of tests the firm picks itself. The useful check this week is a small set of the team's own tasks run against the new model. That shows which of those three differences changes the work.

Source

Introducing Claude Fable 5.2 and Claude Mythos 5.2 (Anthropic)

AINews recap (Latent Space)