Anthropic Released One Model Under Two Names. Here Is What You Actually Get.
Fable 5.2 is the version anyone can use. The real story is what a thousand copies of it finished in 30 hours, and what that run does not prove.
Imagine hiring a thousand accountants on Monday morning. You hand them one of the most famous proofs in mathematics and ask them to rewrite it so that a machine can check every single step. There is no manager and the brief is one page. By Tuesday afternoon they are done.
That is roughly what Anthropic says happened inside its lab. The model that did it shipped this week under two different names. You are allowed to use one of them.
What happened
On November 9, Anthropic released Claude Fable 5.2 and Claude Mythos 5.2. The announcement says it in one line: “Claude Fable and Claude Mythos are the same underlying model.”
The difference is a gate. Fable carries safeguards for cybersecurity and biology, “which allows us to make it available to all users.” Mythos is the same model without those safeguards. Organizations that work in cybersecurity or biomedical research have to apply for it through Anthropic’s verification programs.
The rest of the release is about price and speed:
- $8 per million input tokens and $40 per million output tokens, which Anthropic calls “a lower price” than the previous version.
- Output up to 30% faster than Fable 5.1.
- An optional fast mode, up to 2.5 times quicker, at $16 and $80.
- About 55% fewer blocked requests than Fable 5.1. When a request is blocked, an older model answers instead: Opus 4.8 for cybersecurity, Opus 5 for biology.
- For subscribers, Fable counts for up to 50% of weekly usage limits on Pro plans and on standard Team seats.
The 30-hour run, with the fine print
The headline result is a formalization of Fermat’s Last Theorem. Formalization, in Anthropic’s words, is “the process of translating a mathematical proof into a language that a computer can check step by step.”
The theorem itself is not news. Andrew Wiles published the first correct proof in May 1995. What is new is the machine-checkable version, and how fast it now gets built.
Anthropic had already done this once. In that first run, dozens of agents on an internal research model worked for 11 days, with occasional high-level instructions from a researcher. They wrote 13 million lines of Lean, the language the checker reads, and used about six billion output tokens.
This time, 1,000 Fable 5.2 agents did the full job in 30 hours, “with minimal human intervention.” According to the announcement, the agents started from minimal prompts, built their own ways to coordinate and used no mathematical references. Their one aid was a list of intermediate theorem statements to score progress. They needed one-third as many action tokens, and the proof came out seven times shorter.
Then Anthropic ran the same setup with Opus 5.5, the model it calls its “strongest daily driver.” Same thousand agents, same task. It “completed only 26% of the formalization in 18 hours, and barely made any more progress after 30.”
Now the fine print, all of it from Anthropic’s own text:
- It was a single run per model. One race, not a championship.
- Both teams were stopped at 18 hours and resumed from their last save with 12 more hours.
- The numbers are Anthropic’s own. We found no independent replication.
- It is not a discovery. The agents rebuilt a proof that already existed.
Why it matters
Most AI news is about answers. This one is about finishing.
The two teams had the same size, the same instructions and the same clock. Only the model changed. One team reached the end. The other stalled a quarter of the way in and stayed there. Anyone who has managed a long project knows that picture. Starting is cheap. Arriving is the hard part.
Get Alius’s stories in your inbox
Join Medium for free to get updates from this writer.
That is the shift business owners should notice. The useful question is no longer “how smart is the answer.” It is “does the long job get done without someone holding its hand.”
There is a second lesson in the price. Anthropic does not tell you to use Fable for everything. It says the opposite: “For most coding and knowledge tasks, we recommend that users start with Claude Opus 5.5, our strongest daily driver.” Fable and Mythos, it says, are “best for the most ambitious, open-ended technical work.”
Read that as plain advice. The expensive model is for the job that would otherwise not get done at all.
And a run like this is not cheap. Here is our own rough arithmetic, not Anthropic’s figure. The first formalization used about six billion output tokens. If the new run used a third of that, and if those tokens were billed at $40 per million, the bill lands near $80,000. The real number may differ, because action tokens and output tokens are not the same thing. The order of magnitude is the point. A thousand agents for 30 hours is a project budget, not a subscription.
What else is in the release
Three items deserve a line, because they are closer to real discovery than the Fermat run.
Physics. Anthropic says Fable 5.2 “solved two open problems in theoretical physics,” both about black holes. One was a calculation laid out in 2003 that was “too laborious to complete.” The results are in preprints. We have not seen independent review yet.
Hardware. From a one-paragraph prompt, the model designed three working devices with their own circuit boards: a home voice-assistant hub, a USB speakerphone and a handheld virtual pet. Each board was ready for manufacturing in under a day. “All three devices worked without needing any electrical modifications.”
Software. One engineer asked it for a prototype operating system. The model coordinated over 1,000 subagents for eight days and produced a system that boots on a laptop.
Who gains, and who hits the wall
Anyone with a long, boring, well-defined job. Moving an old program to a new language. Checking a year of documents against a rule. Rebuilding a messy database. Anthropic reports that agents migrated one of its own production services from Python to Rust, and that the service now “starts about 26 times faster.”
Security teams that get verified. According to the announcement, one security company found that one in six confirmed bugs found by Mythos 5.2 were missed by every other frontier model it tested.
Everyone else in security or biology hits the filters. The announcement is direct: Fable 5.2’s safeguards “will block most cybersecurity-related tasks.” If that is your field, the public model will refuse more than you expect.
What to do this week
1. Write down the long jobs you never start. The migration, the clean-up, the audit that would take a person three weeks. That list is where this class of model changes the math. Email drafts are not on it. 2. Keep daily work on the cheaper model. Follow the seller’s own advice. Move up only when the job is big and open-ended. 3. Ask whoever runs your software three questions. Which model touches our data? On which plan, and under which terms? When a request gets blocked, which older model answers instead? 4. If you work in security or biomedical research, apply for verification. It is the supported path. Fighting the filters is not. 5. Run one test with a finish line. Pick a task where “done” is obvious. Count three things: did it finish, what did it cost, and how much did a person have to fix.
What we expect next
These are forecasts, not facts. For each one we say what would prove us wrong.
Access by verification becomes normal. One model, different gates for different customers. The benchmark table in the announcement already lists a rival variant called GPT-6 Cyber. We expect every major lab to sell its strongest model this way within a year. We are wrong if the labs return to one public version for everyone.
Cost per finished job replaces cost per token. Anthropic’s own charts now report “cost per task.” Buyers will follow. We are wrong if price pages a year from now still lead with tokens alone.
The thousand-agent demo becomes a marketing format. Expect every lab to publish one. The ones worth trusting will come with repeated runs and outside replication. We are wrong if independent teams reproduce this 30-hour result quickly and the gap between models holds.
Machine-checked work spreads beyond mathematics. Anthropic notes that formal checking “matters even more when AI models generate results that are hard for humans to verify.” Software is the obvious next place. We are wrong if formal verification is still a research curiosity a year from now.
The big picture
The model did not prove Fermat’s Last Theorem. A person did that, three decades ago, after years of work and one nearly fatal gap.
What the model did is less romantic and more useful. It took a job that is enormous, tedious and exact, and it carried it to the end with almost no supervision. Kevin Buzzard, the mathematician who leads the human project to formalize the same proof, reviewed the first machine version. He told Anthropic that it “proves Fermat’s Last Theorem with no assumptions other than the axioms of mathematics.”
Most companies do not have a theorem to formalize. They have the software nobody dares to touch and the archive nobody has time to check.
One team finished and one did not. Same size, same brief, same hours. That is the sentence to keep.
ALIUS is an applied AI lab. We build and run production systems for companies, and we read the AI news so our clients can act on the few stories that matter. Follow for one story a day, checked against the original sources.
Sources: Anthropic, “Introducing Claude Fable 5.2 and Claude Mythos 5.2,” November 9, 2026; Anthropic, “Formalizing Fermat’s Last Theorem”; Decrypt, “AI Just Solved a 350-Year-Old Math Problem By Writing the Longest Proof Ever.”
