Claude Fable and Claude Mythos: everything you need to know

Published on

This week, Anthropic finally released the Claude Mythos 5 and Claude Fable 5 models, and my first impression is that we finally got to see the dragon.

Except the dragon came behind a pane of glass, with a safety warning, fire blockers and a Mario Kart banana peel.

Claude Mythos and Fable cover

Let’s take it piece by piece. According to Anthropic, Mythos 5 and Fable 5 are the same model. The only difference is that Fable has a bunch of guardrails and filters that Mythos doesn’t. That’s exactly why Fable is available to everyone, while Mythos is only for 150 partner companies in the Glasswing project.

Both seem to be a distilled version of what Mythos Preview was. That would explain why they cost US$ 50 per 1M output tokens, and not the absurd US$ 125 from before. Still expensive, but less than half.

Back to Fable, which really is an exceptional model for coding tasks. On the SWE-bench Pro benchmark, Mythos Preview had already reached 77.8%. Now Fable 5/Mythos 5 went up to 80.3%. Not a gigantic jump over the preview, but a huge one over Opus 4.8 and GPT-5.5 (69.2% and 58.6%, respectively). On harder programming benchmarks, like FrontierCode, the gap also looks big.

Fable 5 seems to have been optimized for long, deep, agentic tasks, like investigating a repository, finding bugs, running audits, fixing a bunch of problems and staying coherent for hours.

And that matches the real-world reports:

  • Sérgio Lopes here at Alura used it to review an entire repository, and the model found 68 security issues, ran for hours, ate through usage limits, but in the end produced a flawless pull request.

  • An engineer on a friend’s team one-shotted a computer vision problem he’d been stuck on for 12 days with Opus.

But Fable came with some serious problems.

For biology, chemistry and cybersecurity tasks, it comes with plenty of blocks. According to Anthropic, to prevent its use in creating biological weapons and hacks. In those cases, the question gets routed to Opus 4.8 and you’re told about it. However, their classifier came out pretty busted from the factory, full of false positives, and it’s routing even silly questions like “What is a mitochondrion?” to Opus.

To me, the most serious case was/is AI research. A small note, buried in the model’s system card, says that requests related to model training, AI pipelines, Machine Learning, etc. would be treated differently. Instead of simply saying “I can’t help with that” and downgrading to Opus, it said the answer could be altered so as not to help the research move forward. In other words: the model could sabotage you.

Their excuse is that they want to “prevent foreign entities not committed to safety and the good of humanity from being able to build a model as powerful as this one”. Uh-huh.

That is TERRIBLE, not to say absurd. If you’ve read or watched “The Three-Body Problem”, it’s basically the logic of the sophons. The system doesn’t block you directly. It undermines your ability to trust your own experiment. And when that happens in a model you’re paying (a LOT) to use as an intellectual tool, the problem is huge.

And the backlash was huge. The entire AI ecosystem publicly criticized the decision, and it made so much noise that Anthropic said they’ll now at least warn you when it happens. Still terrible, but less terrible.

Another piece of bad news is that they’ve now abandoned Zero Data Retention (your data isn’t stored for anything), and moved to 30-Day Retention (they’ll keep your data on their servers for up to 30 days). According to them, that data won’t be used to train models, only for internal research, abuse evaluation, benchmarks and monitoring attempts at dangerous use. Still, that can cause serious problems when we’re talking about data governance in companies.

Tell your CTO, and be careful.

One more detail is that Fable 5 is temporarily available to Pro and Max plan users, but only until June 22. After that, the initial idea is for access to be restricted to the API, which will be a lot more expensive for users. They said this is temporary, and if they manage to spread the load well across the GPUs and usage doesn’t completely explode, they may make it available again on the monthly plans. We’ll see.

Anyway, Claude Fable 5 really is an excellent model, the best model available today for heavy agentic programming. Anthropic really did manage to make a big leap.

But it comes with lots of asterisks. It’s powerful, but expensive. Impressive, but limited. Safe, but maybe way too fussy about safety. Transparent about some things, opaque about others. Capable of solving tasks other models can’t, but also capable of handing you a different model instead.

And, above all, it puts an important question on the table: when we evaluate a model, are we measuring the capability of the brain or the behavior of the final product?

Because starting with this release, those two things are no longer the same. The model might know, but the system might not let it. The benchmark might measure one thing, but the API might deliver another. And the user is stuck in the middle, trying to figure out whether they’re talking to the dragon or to the dragon’s authorized projection.

My recommendation for now is simple: use Fable 5 if you have access, especially before June 22, for really hard coding, auditing and investigation tasks. But pay attention to the downgrade warnings, the limits, the cost and the data retention policy.

Fable 5 is the first step of the next level of AI models, but it’s also a reminder that the next level doesn’t come with just more intelligence. It comes with more politics, more control, more asterisks and, apparently, more fine print than a medicine leaflet.

And instead of just making Anthropic shine brighter, it created a huge asterisk of distrust about the company among everyone in the AI market and research community.