Bits on Bots

Anthropic shipped one model twice

10 Nov 2026 · News

Claude Fable 5.2 and Claude Mythos 5.2 are the same model. What separates them is a check of who you are.

The most capable AI now ships with an ID desk in front of it, and that desk is the part worth watching. On Monday Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, and its own launch post says they "are the same underlying model." Fable is open to everyone, with filters on security and biology work. Mythos has looser filters and goes only to organizations that pass a verification program. Both run the same weights. What differs is the paperwork you file to get in.

What changes for a regular user

Fable's filters block "penetration testing, exploit generation, and binary-based vulnerability" work, plus most research biology. When a filter blocks a question, an older model answers it instead. Opus 4.8 takes the security questions and Opus 5 takes the biology ones. Anthropic says the filters now trip about 55% less often than they did on Fable 5.1, and that the safeguards are as robust as before.

So for most people, nothing is lost. Ask Fable to fix a bug and you get Fable. Ask it to write an exploit and you get a lesser model's answer. If you need the full model for that work, you apply to the Cyber Verification Program or its life sciences twin, and Anthropic decides whether you are who you say you are.

The same post explains why the check matters. Anthropic says illicit distillation, training a rival model on Claude's answers, runs "on an industrial scale, often involving thousands of fake Claude accounts," and that its new protections cut the data those accounts harvest by 90% to 99%. Fake accounts are an identity problem.

OpenAI already works this way

This is not one lab's quirk. OpenAI's Daybreak program gives verified defenders "reduced refusals" on the same mainline models everyone else uses, and its top tier unlocks GPT-6 Cyber only after "stronger verification." The GPT-6 Cyber system card, published 5 November, lists identity checks, single sign-on, multifactor login and IP allowlists as conditions of access. At both of the biggest labs, the most dangerous tier is now a product you have to qualify for.

Chips went first

Hardware got here earlier, and its record is not encouraging. The US Commerce Department reviewed and approved dozens of Chinese firms to buy or handle Nvidia's H200 chips. One of them, an assembler called New Matrix, lists an office in Beijing that reporters found empty, with wires hanging from the ceiling, and its owners tie back to a private equity fund on the US entity list, per The New York Times. It is the second approved H200 handler with a seemingly deserted address. The review approved the company on its filings, and the office behind them was empty.

Nvidia's own response is another identity step. New cloud partners in Southeast Asia must now enroll in its Fleet Intelligence software, which can record where each accelerator sits, Bloomberg reported. Nvidia does not automatically receive that location data.

Why the desk is the weak point

A model filter judges every question, so fooling it takes work on each request. Anthropic's 55% drop in false blocks suggests its filters are getting more precise. A form that asks who you are gets judged once. A shell company can fill it out. So can a front for a blacklisted fund, as the H200 licenses show. When access depends on identity, attackers will put their effort into faking identity.

That doesn't make tiered access a bad idea. Giving defenders the full model while filtering everyone else is probably the most workable deal on offer. But the safety claim now rests on a vetting process the public can't see.

Then and now

In October we wrote about a bank defending against Mythos and about an agent that wouldn't take no. Both treated the risk as something the model does. Monday's launch puts more of the control in the customer list.

Our call

By 30 June 2027, Anthropic, OpenAI or Google will publicly say it removed an organization from a verified-access program, such as Mythos access or Daybreak, for misrepresenting who it is or passing access to someone else. We will be wrong if no such removal is disclosed by then.

#models #news