Anthropic won't sell you its best model

On Tuesday, Anthropic released the most capable model it has ever built and kept it out of public hands. The uncapped version, Claude Mythos 5, is going only to a short roster of corporate partners, a group the company describes as "select biology researchers," and, Anthropic says, the United States government. Everyone else gets Claude Fable 5, which runs on the same underlying weights but ships with what the company calls guardrails. Ask Fable 5 about cybersecurity, biology, or chemistry and it demurs, routing the request to the less capable Claude Opus 4.8. Try to use its outputs to train a smaller model, a practice known as distillation, and the request is routed to Opus 4.8 as well. Dianne Penn, Anthropic's head of product management for research, concedes the filter is built to err wide, catching benign queries alongside the dangerous ones.

For the median buyer, this is a curious thing to pay up for. Fable 5 costs developers $10 per million tokens of input and $50 per million of output, twice the rate of Anthropic's standard public tier, and by the company's own early data the rerouting fires in fewer than 5% of sessions. The premium, then, is not the cost of a hobbled product; most of the time the buyer is getting the full model. What the buyer cannot get, at any price, is the same model with the walls down. Most write-ups will file this under responsible AI and move on.