2 ms·

I'm sure we'll have true test-time-learning soon (<5 years)but it will be more expensive. Alphaproof (for Deepmind's IMO attempt) already has this.