The fifteen-day problem

The most substantive objection to the claim that Kimi K3 was built by distilling Anthropic’s Fable is not political. It is arithmetic.

The two dates

Fable 5 returned to public availability on 1 July 2026.
Moonshot launched Kimi K3 on 16 July 2026.

That leaves fifteen days. For distillation from Fable to explain K3’s general capability, essentially the entire pipeline would have to fit inside that window: large-scale elicitation of teacher responses, construction and cleaning of a training corpus, a full training run, evaluation, safety work, and release engineering.

Why that is a hard claim

Transferring frontier capability is not a matter of a few thousand examples. It requires query volume on the order of millions of exchanges, the Anthropic complaint six weeks earlier described 28.8 million exchanges over 44 days as the scale of a serious campaign. Then the resulting corpus has to be trained on, and a model of K3’s reported size is not trained in a weekend.

None of this makes the accusation impossible. A model can be substantially complete before a final round of distillation is applied, and a fifteen-day window could in principle accommodate a targeted top-up rather than a from-scratch transfer. But that is a materially weaker claim than the one that was made, and it would need to be stated and evidenced as such.

What has not been published

As of this writing, neither Anthropic nor the White House has published evidence directly connecting the alleged extraction activity to Fable 5 or to Kimi K3. The public case consists of traffic analysis showing that large-scale querying occurred, and a model that arrived sooner and cheaper than expected. The step joining those two things has not been shown.

The underlying problem

This is not a failure of diligence by any particular party. Attributing a trained model to a specific teacher is an unsolved technical problem. Behavioural fingerprints are contaminated because models train on text generated by other models. Weight-space forensics do not survive ordinary retraining. Watermarks degrade under paraphrase.

Until attribution becomes reliable, disputes of this kind will continue to be resolved by politics rather than by proof, which is precisely what happened here.