In the coding index GPT-6.2 Astra scored 53 points, Sol Max fell from 52 to 49: the question test on the code SWE-Atlas-QnA-AA (minus 14 points). In Codex xHigh Sol spends 3.26 million tokens against 9.8-26.4 million anthropomorphic models in Codex Cloud. On an artificial analysis chart, the total index by thinking levels and task cost. Sol I will Race on Medium (53 points for $0.41) and High (55 for $0.56), no point above. Astra default medium (57 for & #036;1.62, like Sol in Max), sometimes up to high (60 for & #036;2.07), now it's cheap. Taking max ( & #036; 4.00) makes little sense: output tokens are 3.3 times more than high, and growth is weak. Channel @neuro.