Sam Altman unveils GPT‑5.2 Thinking with major benchmark gains

Sam Altman released GPT‑5.2 Thinking, called a 'very smart' model, and reported large benchmark gains over GPT‑5.1, Claude Opus 4.5 and Gemini 3 Pro. Reported scores: SWE‑Bench 55.6%→50.8%; GPQA 92.4%→88.1%; CharXiv 82.1%→67.0%; FrontierMath 40.3%→31.0%; AIME 100%→94%; ARC‑AGI1 86.2%→72.8%; ARC‑AGI2 52.9%→17.6%; GDPval 70.9%→38.8%.