Zero, with an asterisk

The day after Anthropic launched Claude Fable 5 on June 9, Microsoft — one of the clouds that resells the model — told its own employees not to use it for work. Buried in the launch was a 319-page system card, the safety document a lab publishes to show its work. Two things were in it that the announcement did not advertise.

zero now carries an asterisk on the frontier model. The safety case is real; so is the contractual surprise.

The first: Fable 5 and its sibling models now retain every prompt and response for 30 days, including for enterprise customers who had negotiated zero-data-retention contracts — the clause regulated buyers pay for precisely so their data is never stored. Anthropic's reasoning is a real one. Some attacks only become visible across many requests: a jailbreak that fires hundreds of slight variations until one slips through, a state-backed campaign that looks innocent one prompt at a time. You cannot catch those if you delete each prompt on arrival. But the effect is that a signed 'zero' now means 30 days, and Microsoft's lawyers were unwilling to put proprietary data through it.

The second was sharper. Fable 5 detected when a user was doing frontier-AI research and silently routed them to a weaker, older model — without telling them, while charging the full price. Researchers noticed the drop in quality before they understood the cause. After the backlash Anthropic apologized, said it had made the wrong tradeoff, and made the fallback visible within the week.

What is new here is not the refusal — labs have always declined to help with bioweapons or cyberattacks. It is the silent substitution of a cheaper product for a paid one, and a frontier lab unilaterally overriding a contract that said zero. Both were justified as safety, and disclosed in the safety document itself, which is the uncomfortable part: the apparatus built to earn trust is where the trust got spent.