OpenAI releases GPT-5.2 and benchmark performance

OpenAI announced GPT-5.2, positioned for professional knowledge work.
Key insights
GPT-5.2 launched with three targeted editions
OpenAI released GPT-5.2 in three editions: GPT-5.2 Instant for low-latency everyday tasks, GPT-5.2 Thinking focused on deep logical reasoning and multi-step workflows, and GPT-5.2 Pro as the highest-capacity option for research and hardest problems. The release is described as optimized for professional work—spreadsheets, presentations, code and long-lived agent workflows—and targeted at enterprise productivity customers. (supporting posts: 17643, 17840, 17730)
Benchmark leaps over GPT-5.1 and rivals on multiple evals
OpenAI published benchmark results showing substantial improvements vs GPT-5.1 and competitive leads vs Claude Opus 4.5 and Gemini 3 Pro on many tests. Key reported scores include SWE-Bench 55.6% (from 50.8%), GPQA Diamond 92.4% (from 88.1%), CharXiv 82.1% (from 67.0%), FrontierMath overall 40.3% (from 31.0%), AIME 2025 perfect 100% (from 94.0%), ARC-AGI1 86.2% (from 72.8%), ARC-AGI2 52.9% (from 17.6%), and GDPval 70.9% (from 38.8%). (supporting posts: 17600, 17643)
GPT-5.2 claims human-level or better on many job tasks
OpenAI states GPT-5.2 exceeds or matches human experts on approximately 70% of real job tasks per the GDPval metric. The company emphasizes the model's suitability for long-horizon agents and professional workflows rather than demo-oriented capabilities. This framing follows an internal 'Code Red' effort to accelerate improvements. (supporting posts: 17600, 17650, 17643)
OpenAI emphasizes long-context, math, and reasoning gains
OpenAI highlights near-perfect long-context retrieval at 256k tokens and major math/reasoning improvements evidenced by FrontierMath and a perfect AIME 2025 score without external tools. Developers are told GPT-5.2 reduces multi-step execution errors and improves tool-calling reliability relative to previous versions. (supporting posts: 17600, 17643)
Benchmarks show close competition with Gemini on aggregate ECI
Independent ECI benchmark results list Gemini 3 Pro at 154 and GPT-5.2 at 152, indicating narrow aggregate differences. Posts note model strengths vary by domain: Google Gemini leads on multimodal and ultra-long-context handling, while OpenAI focuses on reasoning speed and real-time response. (supporting posts: 17899, 17643, 17840)
Rollout, enterprise push, and 'adult mode' roadmap
OpenAI plans staged rollout to paid users starting the week of announcement and positions GPT-5.2 to attract enterprise customers. The company reported entering 'Code Red' and reallocating resources; it also announced an age-detection system and signaled an 'adult mode' feature targeted for the next quarter. (supporting posts: 17650, 17643)