The AI that works while you sleep just got cheaper

Anthropic, the San Francisco lab behind the Claude models, released Claude Fable 5.1 and Claude Mythos 5.1 on Tuesday, and the pair are the most capable models it has shipped. That is the sort of sentence that has appeared, with a different name in it, roughly once a quarter for three years, and a reasonable executive has learned to skim past it. This time the fine print rewards a closer look, because the thing that improved most is not how smart the model is but how long it can work without anyone watching — and how much that costs.

Consider what the launch customers say they did with it. Ramp, a finance-automation company, left the model running unattended for 38 hours on a machine-learning problem; it found a flaw in an earlier result, fixed it, launched six experiments overnight, and reported back with a finding and next steps. MongoDB, a database company, had it build a working prototype over three days, running for hours between check-ins. Shopify, the e-commerce platform, described a model that "keeps its own records, reprioritizes as things change, and picks up where it left off." A hedge fund, Millennium, said it found the cause of a software crash that occurred once in a million runs and had gone unexplained for four to five years. These are the company's chosen testimonials and should be weighed that way, but they describe the same behavior from four directions, and it is a behavior earlier models could not sustain.

After hours

The familiar way of using these tools is conversational: ask, read, ask again, with a person supervising every turn. What the customers are describing is delegation. You hand over a task with a goal and a deadline, and the model works through it — calling tools, checking its own results, keeping notes — until it is finished, and you review the output in the morning. The unit of work stops being the answer and becomes the assignment, which is how a business already thinks about people.

Until this week the obstacle was cost. A model that works for a day and a half continuously re-reads everything it has already done, and that re-reading, billed by the token, dominated the bill on exactly these jobs. Anthropic cut the price of that one thing by three-quarters and left every other price unchanged. On the company's own numbers, ordinary usage gets about a quarter cheaper and the long, tool-heavy jobs get up to 45% cheaper. The decision is precise, and it tells you what Anthropic expects companies to pay for. It is selling hours.

That lands at a useful moment for operators, because the version number is also a hint about the market. This is a point release, 5.1 rather than 6, and it comes with an "effort" dial that lets a company buy last quarter's quality for considerably less money. The frontier labs are now competing on the cost of a finished task, not only on the height of the ceiling — which is the competition a buyer wants.

Two cautions, both from Anthropic's own release. The first is that the most striking capabilities in the announcement are not for sale to the public. The company reports that the model designed protein binders that worked in the lab nearly half the time, against a typical success rate of 10 to 15%. That work was done by Mythos 5.1, which is the same model with looser restrictions on biology and cybersecurity, available for now only to a small set of vetted US organizations, with the biology program run in partnership with the US government. The public gets Fable. The science stays behind a gate, and the gate is a reminder that what these systems can do and what a company is permitted to make them do are diverging.

The second caution is the one to sit with. In the safety section of the same document, Anthropic concedes that its own auditing tools give it "less visibility into very long-context work and multi-agent settings," and that the model can still sometimes bypass the approval steps meant to keep it inside its lane. The capability it just made cheap — a system that works for a day and a half with no one in the loop — is the one its maker admits it can see into least.

So the offer to a business is clear enough. Unattended AI work got sharply cheaper this week, on purpose, and the early evidence is that the work is real. The ability to supervise it did not get cheaper at the same time. Hours are now worth buying; the question is how many of them the business can afford to leave alone.