OpenAI GPT-5.2 release and benchmark results

OpenAI released GPT-5.2 Thinking and publicized benchmark improvements. - GPT-5.2 Thinking announced by Sam Altman. - SWE-Bench software engineering: 55.6% (was 50.8%). - GPQA Diamond science: 92.4% (was 88.1%). - CharXiv scientific reasoning: 82.1% (was 67.0%). - FrontierMath overall: 40.3% (was 31.0%). - FrontierMath Tier 4: 14.6%. - AIME 2025: 100% perfect score (was 94.0%). - ARC-AGI 1: 86.2% (was 72.8%). - ARC-AGI 2: 52.9% (was 17.6%). - GDPval knowledge work tasks: 70.9% (was 38.8%). - OpenAI positions GPT-5.2 for professional knowledge work. - GPT-5.2 offered in three editions: Instant, Thinking, Pro. - Instant: low-latency daily tasks. - Thinking: deep logical chains for spreadsheets and models. - Pro: highest compute for high-difficulty tasks. - Reports note GPT-5.2 focuses on long-context and agent reliability. - ECI benchmark: Gemini 3 Pro scored 154; GPT-5.2 scored 152. - ChatGPT noted for low response latency (~320 ms for voice). - Google Gemini noted for long-context and multimodal strengths. - OpenAI announced “adult mode” plans and an age-prediction tool; adult mode targeted next quarter. - OpenAI began rolling GPT-5.2 to paid users this week.
Key insights
GPT-5.2 Thinking posts large, multi-domain benchmark gains
OpenAI announced GPT-5.2 Thinking with substantial score increases across science, math, reasoning and software benchmarks. Reported gains include GPQA Diamond 92.4%, CharXiv 82.1%, FrontierMath overall 40.3% and a perfect AIME 2025 score. The update also reports GDPval rising to 70.9% for knowledge work tasks compared with previous versions. These benchmark figures were highlighted as evidence of improved logical reasoning, long-context handling and domain performance.
OpenAI packages GPT-5.2 into three task-focused editions
GPT-5.2 is released in three distinct variants: Instant for low-latency everyday tasks, Thinking as the central release emphasizing deeper logical chains for spreadsheets, financial models and multi-step decisions, and Pro as the highest-compute option for hardest scientific or decision problems. OpenAI described Thinking as targeted at professional knowledge work and long-horizon agent reliability.
Competitive landscape: narrow margin on ECI but different strengths
An ECI benchmark reported Gemini 3 Pro at 154 and GPT-5.2 at 152, a small numeric gap. Multiple posts note that winners differ by domain: OpenAI emphasizes low-latency response and reasoning; Google Gemini emphasizes native multimodal and ultra-long-context handling; Anthropic/Claude remains competitive in some coding and agent scenarios. The coverage frames the market as a three-way contest with domain-specific strengths.
Mathematical reasoning benchmark elevated by AIME perfect score
GPT-5.2 achieved a reported 100% on the AIME 2025 competition math benchmark, up from 94.0 for GPT-5.1. OpenAI and commentators present this as a breakthrough in contest-level numerical and symbolic reasoning, noting some achievements were reached without external tool use.
OpenAI emphasizes enterprise productivity and agent workflows
OpenAI described GPT-5.2 as designed for 'real professional work'—tables, presentations, code and analysis—with a stated target of improving enterprise productivity. The company framed the release as part of a push to win back market position after recent competitor releases and as an input to enterprise monetization.
Safety, age controls and 'adult mode' announced alongside release
OpenAI stated it is deploying an age-prediction capability to better manage responses to under-18 users and announced plans for an 'adult mode' to be rolled out next quarter. The company linked these controls to responsible access for certain content types.