34 hours of agent time, about 40 minutes on the actual research question

One researcher ran the same self-distillation task (training a model on its own outputs) through two model generations on identical hardware. Opus 4.8 got two days across three sessions. Fable 5 worked 34 hours straight.

Opus racked up six errors in the objective function, and its self-check caught none of them. In one episode it fired a 97-prompt check at a live server and killed a 1.4-hour run at step 43.

Fable 5 wrote a plan before every run and caught a bug at the design stage. But across all 34 hours it never once checked its work against the goal, and the research question itself got roughly 40 minutes.

All eight failure classes showed up again. In its post When AI builds itself, Anthropic said Claude wrote more than 80% of the code in its repository.