OpenAI releases GPT-5.2 and benchmark results

OpenAI released GPT-5.2, emphasizing professional knowledge work and long-running agents. - GPT-5.2 includes Instant, Thinking, and Pro variants. - GPT-5.2 Thinking posts large benchmark gains over 5.1. - SWE-Bench software engineering: 55.6% (from 50.8%). - GPQA Diamond science: 92.4% (from 88.1%). - CharXiv scientific reasoning: 82.1% (from 67.0%). - FrontierMath overall: 40.3% (from 31.0%); Tier 4: 14.6%. - AIME 2025 competition math: 100% (from 94.0%). - ARC-AGI 1: 86.2% (from 72.8%); ARC-AGI 2: 52.9% (from 17.6%). - GDPval knowledge-work metric: 70.9% (from 38.8%). - 256k token context tests near-perfect long-context accuracy. - OpenAI described a “Code Red” internal push preceding release. - GPT-5.2 rollout to paid users begins this week; “adult mode” and age detection planned.
Key insights
GPT-5.2 framed as workplace-grade model with three variants
OpenAI positioned GPT-5.2 as a shift from chat-focused models toward "professional colleague" capabilities by releasing three specialized variants: Instant for fast everyday tasks, Thinking for deep multi-step reasoning and long-horizon agents, and Pro for highest compute scientific or research tasks. The company emphasized reliability for knowledge work and long-cycle agents rather than demo-oriented capabilities, aiming to recover competitive standing versus Gemini and Claude lines.
Benchmark improvements show large gains across reasoning and math
Reported benchmark results for GPT-5.2 Thinking demonstrate notable increases over GPT-5.1: SWE-Bench 55.6% (from 50.8%), GPQA Diamond 92.4% (from 88.1%), CharXiv 82.1% (from 67.0%), FrontierMath 40.3% (from 31.0%), and a perfect 100% on AIME 2025 (from 94.0%). Abstract reasoning scores rose sharply (ARC-AGI 1: 86.2% from 72.8%; ARC-AGI 2: 52.9% from 17.6%), indicating improved multi-step and abstract problem-solving performance.
OpenAI cites GDPval metric and claims high human parity on tasks
OpenAI introduced GDPval as a task-level benchmark for knowledge work and reported GPT-5.2 achieving 70.9% of tasks at or above human-expert performance, up from 38.8% for GPT-5. This metric was presented alongside claims that GPT-5.2 outperforms competitors on key evaluations and is designed to target real workplace tasks rather than demos.
Long-context and agent design highlighted as core advances
GPT-5.2 is reported to handle extremely long contexts (256k tokens) with near-perfect retrieval accuracy, and is explicitly optimized for long-running agents. OpenAI emphasized reduced error rates in multi-step tool and cross-document workflows and described architecture and training work supporting extended reasoning chains without external tool dependence.
Release timing, corporate strategy, and policy features announced
OpenAI began rolling GPT-5.2 to paying users this week and announced further features: age-detection software to gate responses, and an "adult mode" planned for the next quarter. Company leadership referenced a prior "Code Red" emergency directive to focus resources on ChatGPT improvements ahead of this release, and stated commercial goals tied to enterprise customers.