Eight thousand posts on the first day, and the sentence that mattered was about a cache.
Anthropic released Claude Fable 5.1 on 1 September 2026. The sticker prices did not move: ten dollars a million tokens in, fifty out. Cache reads, the reread of work already done, fell from a dollar to twenty-five cents. They say typical jobs run about a quarter cheaper, and the long agentic ones closer to forty-five. That is the product. Not a new number on a board. A model you can afford to leave overnight.
Fable 5.1 and Mythos 5.1 are the same weights. Fable is the public one. Mythos keeps looser safeguards for a short list of cyber and life-science labs. The company is also trying to say no less often to honest work. Cyber classifiers fire about sixty percent less on false alarms, and the model may now hunt for holes in software, though not write the exploit. Biology still gets sent elsewhere unless you are on the trusted list.
The charts are loud in the usual way. On Terminal-Bench-Science 0.1 it scored 52.6 percent, up from 24.7 on Fable 5. Terminal-Bench 4.0: 55.8 against 42. CursorBench 3.2: 73.4 against 70.5. Fine. The story I prefer is smaller. An investment firm called Millennium had a crash in a vendor library, one in a million runs, unexplained for four or five years. Fable 5 missed it. 5.1 disassembled the library, matched the core dump, and named the bug. That is not an intelligence index. That is a person who got to go home.
Simon Willison, who still asks models to draw a pelican on a bicycle, found the other knob. Effort goes low, medium, high, extra-high, max. At max the pelican took fourteen minutes, sixty-six thousand output tokens, and three dollars thirty. It wore a blue hat and put a fish in the basket. At low and medium the model barely seemed to think, and the bird was only a bird. The effort setting is the real price list. Default in Claude Code is high. On the website, medium.
X did what X does. About 8,134 English posts on the first day, not counting reposts, and another 6,093 by the next afternoon. Boris Cherny, who builds Claude Code, called it a pleasure and said he has been using it for everything. Ben Davis ran the cache maths on a month of his own Fable traffic and got about $1,752 back, if the new rate had applied. Then the other half of the street arrived. People on the $200 Max plan said a session died in an hour. Someone wrote that they would refrain from trying Fable 5.1, seven times, like lines on a blackboard. The cheaper reread is an API line. A subscription is a bucket. Mixing the two is how you get a fight in public about a discount you cannot see on the invoice.
Hacker News, for once, talked about sentences. Felix Rieseberg, who works at Anthropic, said 5.1 sounds less like the other Claudes and takes style instructions more seriously. The thread under him is still about Opus 5’s private dialect, the booked half-days you do not get back. If 5.1 really writes like a person, that is the upgrade a lot of people actually wanted, and it will not show on Terminal-Bench.
CodeRabbit timed the reviews. Same known bugs, eighty-seven fewer comments, a hundred and eighty-six fewer nits, six extra minutes on the clock. Precision up a little. Turning the reasoning up did not help. The model is slower when you let it think, and sometimes better for it, and sometimes not.
So: cheaper to leave running, if you pay by the token. Less eager to refuse the honest job. A little more willing to stay on a crash until morning. The quota still empties if you point every helper at max. That was true yesterday. It is still true. The cheaper reread only helps if you were rereading.