SAN FRANCISCO – OpenAI has officially announced the release of GPT-5.2 on the 12th of December 2025 , a major update to its flagship model series designed to master complex, real-world professional tasks. Described in the company’s latest blog post, the release introduces a trio of new models:
- GPT-5.2 Instant,
- GPT 5.2 Thinking,
- GPT-5.2 Pro.
The new models are engineered to “unlock even more economic value” by acting as comprehensive partners in knowledge work. The timing of the release is peculiar, as Google has recently released their Gemini 3 model. Although common for OpenAI to respond quickly, it has in the past raised concerns on the actual quality of said releases.
Focus on professional utility
According to OpenAI, GPT-5.2 represents a significant shift from simple chat interactions to “agentic” workflows. The models are specifically optimized for end-to-end execution of complex tasks, such as creating formatted spreadsheets, building slide presentations, writing and debugging production-grade code, and analyzing long documents without losing coherence.
“GPT-5.2 Thinking is the best model yet for real-world, professional use,” the announcement states. The company highlights that for the first time, an AI model has achieved expert-level performance on GDPval, a new internal benchmark measuring proficiency in well-specified knowledge work across 44 distinct occupations (including sales, accounting, and engineering).
- GDPval Results: GPT-5.2 Thinking reportedly beats or ties top industry professionals in 70.9% of comparisons, a massive leap from the 38.8% score of its predecessor, GPT-5.
Three new models
The release breaks the GPT-5.2 series into three distinct tiers to cater to different user needs:
- GPT-5.2 Thinking: The powerhouse of the series, designed for deep reasoning and complex problem-solving. It utilizes new “agentic tool-calling” capabilities to handle multi-step projects autonomously.
- GPT-5.2 Instant: A faster, streamlined version optimized for low-latency, everyday tasks.
- GPT-5.2 Pro: A high-capacity model balanced for reliable performance across a wide range of enterprise applications.
Record-breaking benchmarks
OpenAI’s announcement details several state-of-the-art achievements for the “Thinking” variant:
- Coding: Scored 55.6% on SWE-Bench Pro (a rigorous software engineering test), setting a new standard for solving real-world GitHub issues in multiple languages.
- Math: Achieved a perfect 100% on the AIME 2025 mathematics competition benchmark.
- Accuracy: The model reportedly hallucinates less, with a 30% reduction in response errors compared to GPT-5.1 Thinking, making it more dependable for critical business analysis.
Benchmark Comparison: GPT-5.2 vs. Gemini 3
The competition between OpenAI and Google now focuses heavily on reasoning and professional utility (agentic workflows). OpenAI claims leadership specifically in applied work tasks (GDPval), while Gemini 3 remains very strong in pure scientific benchmarks.
| Benchmark | Category | GPT-5.2 (Thinking/Pro) | Google Gemini 3 (Pro/Deep Think) | Winner |
| GDPval | Professional knowledge Work (Win/Tie Rate) | 70.9% (Thinking) / 74.1% (Pro) | 53.3% (Pro) | OpenAI |
| SWE-bench Verified | Software engineering (Bug fixing) | 80.0% | 76.2% | OpenAI |
| AIME 2025 | Math (No tools) | 100% | 95.0% | OpenAI |
| ARC-AGI-2 | Abstract reasoning | 52.9% | 45.1% (Deep Think) | OpenAI |
| GPQA Diamond | Science (PhD-level) | 93.2% (Pro) | 93.8% (Deep Think) | |
| Humanity’s Last Exam | Complex academic questions | 34.5% | 37.5% |
Availability
The new models are rolling out immediately to users on ChatGPT Plus, Pro, and Enterprise plans. OpenAI has also made the models available to developers via the API, allowing for the integration of these new “agentic” capabilities into third-party applications.
The release comes as OpenAI seeks to solidify its lead in the generative AI space, positioning GPT-5.2 not just as a chatbot, but as a “powerful daily partner” capable of performing tangible economic work.
For more information visit the official announcement of GPT 5.2
