XNews

OpenAI GPT-5.2 release and benchmark results

OpenAI GPT-5.2 release and benchmark results

OpenAI announced GPT-5.2 release, focused on professional work and long-range agents. - GPT-5.2 ships in three versions: Instant, Thinking, Pro. - GPT-5.2 Thinking posted large benchmark gains across multiple suites. - 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 advanced math: 40.3% overall (from 31.0%), Tier 4 at 14.6%. - AIME 2025 competition math: 100% (from 94.0%). - ARC-AGI 1 abstract reasoning: 86.2% (from 72.8%). - ARC-AGI 2 abstract reasoning: 52.9% (from 17.6%). - GDPval knowledge-work metric: 70.9% (from 38.8%). - ECI ranking: Gemini 3 Pro scored 154, GPT-5.2 scored 152 in one report. - OpenAI positions GPT-5.2 as focused on 'professional knowledge work' and long-lived agents. - Sam Altman and company messaging referenced an internal 'Code Red' push to recover positioning. - OpenAI said GPT-5.2 will roll out to paid users this week; an "adult mode" and age-detection feature were mentioned, with adult mode planned for Q1 next year. - OpenAI framed GPT-5.2 as optimized for spreadsheets, presentations, code, and long-context reasoning.

Key insights

GPT-5.2 achieves large benchmark improvements across reasoning and math

GPT-5.2 Thinking shows major benchmark gains versus GPT-5.1 across multiple, named evaluations: SWE-Bench rose to 55.6% (from 50.8%), GPQA Diamond to 92.4% (from 88.1%), CharXiv to 82.1% (from 67.0%), FrontierMath overall to 40.3% (from 31.0%) with Tier 4 at 14.6%, and AIME 2025 reached a perfect 100% (from 94.0%). ARC-AGI 1 rose to 86.2% (from 72.8%) and ARC-AGI 2 jumped to 52.9% (from 17.6%). These results were published alongside OpenAI commentary positioning the model for professional tasks and long-lived agents.

Product segmentation: Instant, Thinking, Pro for different workloads

OpenAI released three GPT-5.2 variants: GPT-5.2 Instant for low-latency retrieval and light tasks; GPT-5.2 Thinking focused on deep multi-step reasoning, spreadsheets, and financial modelling; GPT-5.2 Pro as the highest-compute option for difficult scientific and decision problems. The company described Thinking as the core update and Pro as intended for high-difficulty use cases.

GPT-5.2 claims large productivity gains on real-world tasks (GDPval metric)

OpenAI introduced or cited GDPval as a metric for knowledge-work performance; GPT-5.2 reportedly achieves 70.9% on GDPval tasks, up from 38.8% for prior GPT-5, a stated increase indicating higher task-level parity or superiority versus human experts on many job tasks. OpenAI positioned GPT-5.2 toward enterprise productivity and professional workflows (spreadsheets, presentations, code, analysis).

Competitive positioning vs Gemini 3 Pro and Claude Opus 4.5

Public comparisons place Gemini 3 Pro and Claude Opus 4.5 as principal rivals. One public ranking (ECI) showed Gemini 3 Pro at 154 and GPT-5.2 at 152. Reports state GPT-5.2 surpasses GPT-5.1 by large margins on reasoning benchmarks and is explicitly optimized to counter Claude's multi-step agent strengths and Gemini's multimodal advantages.

Rollout, safety and product features: adult mode and age detection

OpenAI announced GPT-5.2 deployment to paid users beginning this week. The company described an upcoming "adult mode" planned for Q1 and said it is deploying an age-prediction feature to gate responses for under-18 users. Corporate communications referenced a 'Code Red' operational posture and infrastructure investment commitments.

Focus on long-context and agent reliability

OpenAI emphasized GPT-5.2's near-perfect performance on ultra-long contexts (tests at 256k tokens reported very high retrieval accuracy), and stated the release is engineered for long-running agents and multi-step tool use, reducing multi-step execution errors and aiming for higher reliability in sustained workflows.

Source posts