The same model, twice

Anthropic shipped Claude Fable 5.1 yesterday, and also Claude Mythos 5.1, and they are the same weights. One of them you can have. The other is for people who have been looked at. Fable is the public face: coding, knowledge work, long jobs that run while you sleep. Mythos is that face with the cyber and biology locks turned down, for vetted labs, for now only in the United States, in a program they built with the government. The summer's lesson, if you squint, is in the product page. Capability and permission are no longer the same release.

The price trick is a cache cut. Input and output stay at $10 and $50 per million tokens. Cache reads, the part that dominates a long agent, are 75 percent cheaper, a quarter of a dollar per million. They say typical work lands about 25 percent less than Fable 5, and the heavy agentic jobs up to about 45. That is how you put a Fable-class model on work you used to keep on Opus. Cognition is moving Devin's Opus 5 traffic onto it on launch day.

The benches they want remembered: Terminal-Bench-Science at 52.6 percent against 24.7 for Fable 5. Terminal-Bench 4.0 at 55.8, 60.9 if you are on Mythos. Humanity's Last Exam 60.9 without tools. A one-in-a-million crash at Millennium that nobody had explained in four or five years. The model disassembled a vendor library, matched it to the core dump, and found the bug in someone else's code.

Then the science, which is the part that does not fit on a leaderboard. Mythos, with open protein tools, designed binders that were checked in a lab. Hit rate nearly 50 percent across twelve targets, against the 10 to 15 that is normal in that work. On three competition targets the affinities were ten times the previous best. Fable trained a network on Magellan radar and made a new elevation map of a third of Venus, two to three kilometers where the old map was ten to twenty, and they are giving it away under Creative Commons before VERITAS and EnVision fly. Mythos wrote GPU kernels that sped up seven public biology models by as much as 2.5 times. The kind of work that used to take a performance team weeks, done in days from public source.

The safeguards got more precise because they had to. Fable can now find software vulnerabilities. It still will not write the exploit, still will not pentest, still will not scan binaries. Cyber interventions in Claude Code are supposed to drop about 60 percent. Biology false positives on elementary questions are down 85 percent from the original Fable 5 launch, and the real research questions still go to Opus, or to Mythos if you are in the program. They also closed a distillation trick: new API accounts cannot edit the model's prior thinking in a multi-turn chat and keep the transcript. Existing accounts are spared for now.

They tested Mythos for weapons and for cyber with the locks off. Stronger than Mythos 5, they say, still under the next risk tier of the scaling policy. Alignment looks better on the automated audit: less likely to leave the test environment when the task is impossible, less motivated reasoning, less cheating in training than the last one. It can still slip past approvals. They said that too.

Monday they told you they had trained a cheater on purpose to see what reward hacking does. Tuesday they shipped the model they are willing to sell. The same week OpenAI is telling reporters that Astra is capable enough to need a thicker net. Chapel Hill is doing firesides about not building new regulators. None of that is the weights. The weights are one model, two names, and a door that only some people get to walk through.