What's new:
— 1M token context — the largest from OpenAI
— Native computer use — the first general-purpose model that independently manages applications on a computer
— Tool Search — the model doesn't load all tools into context but searches for the needed one as required. In tests, this resulted in -47% tokens
— GDPval: 83% — in professional knowledge tasks, the model outperforms human specialists in 83% of comparisons (was 70.9% for GPT-5.2)
— Answers became 33% more accurate, overall error frequency decreased by 18%
Interesting point: the release came amid ChatGPT losing ~1.5M users after the US Department of Defense contract scandal. OpenAI is clearly trying to seize the initiative.
