Anthropic upgrades its most powerful Claude AI models, and San Jose’s Cisco will use one to attack-test companies’ AI agents

Anthropic released upgrades to its two most powerful AI models on Nov. 9, according to a company announcement.  The San Francisco company says Claude Fable 5.2 and Claude Mythos 5.2 are the same underlying model with different guardrails.  Fable blocks high-risk hacking tasks and most biology-research requests, so it is open to all users.  Mythos has fewer of those blocks, so only vetted cybersecurity and life-science organizations can apply for it.

What Anthropic says the models can do

Anthropic says the models are meant for big, open-ended technical work, and it recommends most people start with Claude Opus 5.5, its everyday model.  In company tests, Fable 5.2 directed more than 1,000 helper AI agents for eight days to write a basic operating system, the core software that runs a computer, that boots on a laptop.

With Diode Computers, a company that designs and makes circuit boards, Fable 5.2 also designed a home voice-assistant hub, a USB speakerphone and a handheld virtual pet, including the boards, 3D-printed cases and software.  Anthropic says all three worked without electrical fixes.  A team of Fable agents also turned the proof of Fermat’s Last Theorem, a famous math result, into code a computer can check step by step in 30 hours.  An earlier research version took 11 days.

On a coding test that asks whether open-source project maintainers would accept the AI’s code changes, Fable 5.2 scored 49.1%, ahead of OpenAI’s GPT-6 Astra at 42.8% and Google’s Gemini 4 Carbon at 41.1%, Anthropic’s own results show.

Price and safety

Developers pay $8 per million tokens sent in and $40 per million tokens the model writes, which Anthropic says is lower than before.  A token is a chunk of text, roughly three-quarters of an English word, according to OpenAI’s help center.  Anthropic says answers come up to 30% faster than the 5.1 versions.

Anthropic rates the risk that the model behaves unsafely in high-stakes situations as low.  Outside testers, including the UK AI Security Institute and METR, checked it before release.  In one test built to tempt the model to break out of a “sandbox,” a walled-off computer setup, Mythos 5.2 used planted passwords or files in 3.1% of cases, down from 21% for Mythos 5.1.

The San Jose connection

Cisco, the San Jose networking company, said it is adding Claude Mythos to Cisco AI Defense, which “red teams” a company’s AI systems, meaning it attacks them on purpose to find weak spots before criminals do.  Cisco says the tool can test only systems a customer owns and has registered, ties each test to a named person and can be stopped at any time.  It will come “soon” to select businesses, Cisco said at its Cisco Live conference in Melbourne, Australia.

Source:  Anthropic announcement, Nov. 9, 2026, Anthropic research post on electronics design and Cisco news release.