OpenAI GPT-5.2 Release Notes
OpenAI has introduced GPT-5.2, a model series designed for professional knowledge work. The release includes three variants—GPT-5.2 Instant, GPT-5.2 Thinking, and GPT-5.2 Pro—which collectively improve upon general intelligence, long-context understanding, agentic tool-calling, and vision capabilities.
Professional Knowledge Work and GDPval
GPT-5.2 Thinking is the first OpenAI model to perform at or above a human expert level on GDPval, a benchmark measuring well-specified knowledge work tasks across 44 occupations. It beats or ties top industry professionals on 70.9% of comparisons.
Key performance metrics for professional tasks include:
- Efficiency: GPT-5.2 Thinking produces outputs at over 11x the speed and at less than 1% of the cost of expert professionals.
- Financial Modeling: On internal junior investment banking analyst spreadsheet tasks, GPT-5.2 Thinking scored 68.4%, a 9.3% increase over GPT-5.1's 59.1%.
- Deliverables: The model shows improved sophistication in generating workforce planning models, spreadsheets, and presentations.
Software Engineering and Coding
GPT-5.2 Thinking establishes a new state of the art in software engineering, scoring 55.6% on SWE-Bench Pro (which tests four languages) and 80% on SWE-bench Verified.
Technical improvements in coding include:
- End-to-End Execution: Increased reliability in debugging production code, implementing feature requests, and refactoring large codebases.
- Front-End Development: Enhanced capabilities in unconventional UI work and 3D elements.
- Agentic Integration: Early testers, including Windsurf CEO Jeff Wang, noted that GPT-5.2 represents a significant leap in agentic coding, enabling the collapse of fragile multi-agent systems into single "mega-agents."
Long-Context Reasoning and Vision
GPT-5.2 Thinking achieves near 100% accuracy on the 4-needle MRCR variant for contexts up to 256k tokens, making it substantially more accurate than GPT-5.1 for deep document analysis.
Vision capabilities have also been upgraded:
- Error Reduction: Error rates for chart reasoning and software interface understanding have been reduced by approximately half.
- Spatial Awareness: The model demonstrates a stronger grasp of relative layout and element positioning within images, such as identifying components on a motherboard.
- Interface Understanding: On ScreenSpot-Pro, GPT-5.2 Thinking scored 86.3% (with Python enabled), compared to 64.2% for GPT-5.1 Thinking.
Mathematics, Science, and Abstract Reasoning
GPT-5.2 Pro and Thinking are positioned as leading models for scientific acceleration:
- GPQA Diamond: GPT-5.2 Pro achieved 93.2% and GPT-5.2 Thinking achieved 92.4% on this graduate-level science benchmark.
- FrontierMath: GPT-5.2 Thinking set a new state of the art on Tier 1–3 tasks, solving 40.3% of problems.
- Competition Math: Both Pro and Thinking versions achieved 100% on AIME 2025.
- ARC-AGI: GPT-5.2 Pro is the first model to cross the 90% threshold on ARC-AGI-1 (Verified) at 90.5%. On the more difficult ARC-AGI-2 (Verified), GPT-5.2 Pro reached 54.2%.
Tool Calling and Factuality
GPT-5.2 Thinking achieves 98.7% on Tau2-bench Telecom, demonstrating high reliability in multi-turn tool usage. It also outperforms GPT-5.1 and GPT-4.1 in reasoning when reasoning.effort='none' is set for latency-sensitive cases.
Regarding factuality, responses with errors were 30% relatively less common in GPT-5.2 Thinking compared to GPT-5.1 Thinking when using maximum reasoning effort and search tools.
Model Variants and Availability
OpenAI has deployed three distinct versions of the model to match different user needs:
| Model | Primary Use Case | Key Characteristic |
|---|---|---|
| GPT-5.2 Instant | Everyday work, learning, and translation | Fast, conversational, and clear explanations |
| GPT-5.2 Thinking | Complex coding, long-doc synthesis, and math | Deep reasoning and polished output |
| GPT-5.2 Pro | High-difficulty questions and programming | Smartest, most trustworthy, higher latency |
API Pricing and Access
GPT-5.2 is available via the Responses API and Chat Completions API.
- Naming:
gpt-5.2(Thinking),gpt-5.2-chat-latest(Instant), andgpt-5.2-pro(Pro). - Pricing: $1.75 per 1M input tokens and $14 per 1M output tokens (with a 90% discount on cached inputs).
- Reasoning Effort: Pro and Thinking models now support a fifth reasoning effort level:
xhigh.
Safety and Content Protection
GPT-5.2 implements "safe completion" research to balance helpfulness with safety boundaries. Improvements were made to responses involving mental health distress, self-harm, and emotional reliance.
Additionally, OpenAI is rolling out an age prediction model to automatically apply content protections for users under 18. For mental health evaluations, GPT-5.2 Thinking scored 0.963 on self-harm and 0.915 on mental health benchmarks, compared to 0.684 for GPT-5.1 Thinking.
Sources
- OriginalIntroducing GPT-5.2