◢What's changed from the previous version

The standout is agentic coding. On FrontierCode 2.0, an agentic coding benchmark, Fable 5.2 reaches 49.1% at the highest effort, compared with 32.4% for Fable 5.1 and 45.5% for Opus 5.5, according to Anthropic's table¹. The company says it's the highest score of any released model it has tested.

It also says it costs less for the same result. At high effort, Fable 5.2 scores about the same as Opus 5.5's best on that benchmark (45.3% vs 45.5%) for roughly half the cost per task, and it generates output up to 30% faster than Fable 5.1, according to the announcement¹.

Some of the examples Anthropic¹ gives:

• ▹A team of 1,000 Fable 5.2 agents formalized Fermat's Last Theorem in Lean in 30 hours.
• ▹Fable 5.2 solved two open problems in theoretical physics about black holes.
• ▹With the company Diode, it designed three electronic devices that worked without any changes to the circuits.
• ▹According to Anthropic¹, over 90% of the changes merged into its core codebase are now written by Claude, with human guidance.

◢The cybersecurity part

This is what interests me most about the launch. Fable 5.2 blocks most offensive cybersecurity tasks, such as penetration testing or writing exploits, and when it blocks one it hands the request to an older model, Opus 4.8. According to the announcement¹, in early testing it blocks about 60% fewer cybersecurity requests than Fable 5.1, with safeguards just as robust.

It fits with what I wrote about the attack on Taipei's emergency services with 12 AI agents: the attacker had to switch to open models with no safeguards because the commercial ones refused to attack. This two-version approach, an open one with brakes and one without brakes only for vetted users, is along the same lines.

◢Where it is and what it costs

Fable 5.2 has been available since November 9 in the Claude app, in Claude Code and in the API as claude-fable-5-2, as well as on AWS, Google Cloud and Azure. On the API it costs $8 per million input tokens and $40 per million output tokens, according to the announcement¹.

◢Worth keeping in mind

All the performance figures are Anthropic's, apart from customer quotes, and come from its own testing. The company itself recommends starting with Opus 5.5 for most tasks and moving up to Fable only when needed, because for everyday work the difference isn't always worth it.

If you want your team to know which model to use for what, and how to do it safely, that's part of what I cover in my cybersecurity and AI training.