What's this about?
A user reported evidence that Claude Code is silently A/B testing reduced effort levels — that is, quietly giving some users a less thorough version. It reached 215 points and 190 comments.
What exactly is an "effort level"?
Modern models can spend more or less computation on a given request — thinking longer, checking more, trying more approaches. Effort is that dial. Turning it down saves the provider real money and produces answers that are subtly worse rather than obviously broken, which is exactly why silent adjustment upsets people.
What's the community saying?
HN 215 points, 190 comments
- The central grievance is not the cost-saving but the silence: users paying a fixed price want to know which version they are getting.
- Scepticism about the evidence itself, since perceived model quality is notoriously subject to expectation effects.
- Connection to the wider week: if frontier economics are under pressure per the FT story, quiet effort reduction is exactly the pressure valve you would predict.