OpenAI GPT‑5.2 beats 74% of experts on GDPval – 11× cheaper work

OpenAI finally answered the Devstral and Gemini noise with a workhorse: GPT‑5.2 is live across ChatGPT and the API, and on GDPval it now beats or ties human experts on roughly 71–74% of real knowledge‑work tasks. Those tasks usually take 4–8 hours and $150–$200 of billable time; OpenAI claims 5.2 does them over 11× faster at under 1% of the expert cost. This is the first OpenAI release that feels explicitly aimed at replacing mid‑career desk work, not toy prompts.

Product‑wise, you get three tiers: Instant for everyday chat, Thinking for serious reasoning, and Pro as the slow, heavy hitter. Standard 5.2 runs at $1.75 / $14 per 1M input / output tokens with a 400k context window; Pro jumps to $21 / $168 and lives only behind the Responses API for multi‑minute, high‑effort traces. Benchmarks back the positioning: 55.6% on SWE‑Bench Pro, 92.4% on GPQA Diamond, 100% on AIME 2025, and ARC Prize‑verified 90.5% on ARC‑AGI‑1 with a 390× cost‑efficiency gain over last year’s o3 preview.

The System Card is the other big story: hallucinations drop about 30–40%, deceptive tool use falls from 7.7% to 1.6%, and long‑context retrieval stays near‑perfect out past 128k tokens. It’s clearly a better brain for agents and coders, but not magic—you still need routing, evals, and human checks where mistakes are expensive.