Our bet: in 2027 at least two of three big AI labs still charge extra for thinking — resolves 2027-07-14

Letting a model spend more compute and think longer when it answers (test-time compute) buys better answers, and the labs now price that thinking separately; the question is whether this way of charging lasts.

If it lasts, thinking time is a business of its own for frontier labs; if it is folded back into standard prices, the premium disappears.

One side expects thinking-based pricing to continue; the other expects cheaper small models to catch up and wipe out the premium.

SecondSource’s reading: on 2027-07-14, at least two of Anthropic, OpenAI and Google DeepMind will still list a separate plan on their official pricing pages that bills by inference compute (the computing a model does when it actually answers) or by thinking tokens (the word-like units models bill by), such as extended thinking or a reasoning tier.

Our confidence is 55%. We did not re-check the three labs’ current pricing for this card; that check is left for the due date.

arXiv / Google DeepMind + UC Berkeley (Snell, Lee, Xu, Kumar)
“we can improve the efficiency of test-time compute…”

Check the SKU structure on the three labs' official API pricing pages for a separate test-time compute or reasoning pricing tier.

What would overturn it

If only one lab or none keeps such a tier, the reading that test-time compute commands a pricing premium flips.