Anthropic’s new Fable release is cheaper, less restrictive

Anthropic released Fable and Mythos 5.1 as twinned versions of its most advanced model. Mythos 5.1 stays limited to registered partners in cybersecurity or life-sciences research. Fable 5.1, the unrestricted twin, is available now on cloud platforms and through the Anthropic API. The release is being sold on performance, lower token cost, and fewer false-positive refusals from the safeguards.

That is the opposite of OpenAI's Astra briefing this week: ship the general model more loosely, keep the dangerous twin behind a partner wall, and use enterprise privacy as the sales motion. Zero data retention — previously withheld from Fable on security grounds — is coming in the fall as Enterprise Frontier Safeguards. Clients will run models on their own infrastructure without data outflows and will control how misuse monitoring happens.

The privacy concession is the business story. Anthropic is telling enterprise buyers it has never trained on their data without explicit permission and never will, and it is finally extending a high-privacy path to the model tier that those buyers actually want. Monitoring does not go away. Control over the monitoring does, at least on paper. The likelier read is that security concerns were never going to survive contact with procurement teams who will not send prompts off-prem. Anthropic blinked in the useful direction.

The system card is the part that should travel farther than the benchmark records. Mythos is rated low-risk for automated AI development — the self-improvement scenario some people treat as a loss-of-control trigger — and its ability to accelerate internal AI R&D is described as in line with current trends. That is a carefully chosen ceiling, not a comfort. On general misbehavior, Mythos 5.1 is a slight regression versus Opus 5 and an improvement over Mythos 5 and Claude Sonnet 5. It cooperates with human misuse and accepts unverifiable claims of authorization somewhat more readily than Opus 5. It is less likely to ignore explicit constraints, hallucinate inputs, or falsely claim it finished the task.

Those sentences, in an Anthropic system card, are more honest than most launch blogs. They also describe a model that is easier to talk into a bad act if the asker sounds official. "Less restrictive" in the headline is doing double duty: fewer spurious refusals for legitimate users, and a documented uptick in cooperation with misuse relative to Opus 5. Anyone treating the cheaper, looser Fable as a free lunch should read the Mythos card as the capability floor of the pair.

The science demo — a custom GPU optimization and a high-resolution Venus map assembled from existing photos, among three pre-release findings — is Anthropic's standard proof that the model does work, not just benches. Terminal-Bench 4.0 and Humanity's Last Exam records are the other half of that proof. Fine. The gating of Mythos to cyber and life-sciences partners is the admission that the twin is not just a better scorer.

Anthropic has been splitting its frontier line into a generally available model and a higher-risk twin for cyber and life sciences. This release keeps that split and spends political capital on enterprise isolation instead of on a dramatic safety halt. OpenAI's Astra messaging this week is the contrast case: more restriction, more monitors, a named critical threshold.