OpenAI on Thursday released its answer to Google’s impressive Gemini 3 Pro model–GPT-5.2—and by the looks of some head-to-head benchmark test scores, it looks like a winner. The new model took the highest score on a number of benchmark tests covering coding, math, science, tool use, and vision. (Benchmarks should, of course, be combined with real-world use to tell the whole story. But still . . .)
OpenAI says GPT-5.2, which is a reasoning model, achieved expert-level performance scores on its own GDPval benchmark, which evaluates performance on 44 real professional tasks including things like spreadsheet creation, document drafting, presentation building, and more.
GPT-5.2 topped Gemini 3 Pro on the SWE-Bench Pro benchmark (software engineering tasks) with a score of 55.6% (versus Gemini 3 Pro’s 43.3%). It achieved an 86.2% on the ARC-AGI-1 abstract reasoning benchmark, compared to Gemini 3 Pro’s 75% score. It scored a 92.4% on the GPQA Diamond benchmark (science questions), compared with Gemini 3 Pro’s 91.9% score.
The new model comes in three variants. GPT-5.2 Instant is good for seeking information and how-tos, skill-building and study, and career guidance. GPT-5.2 Thinking is good for harder professional tasks like spreadsheet formatting and slideshow creation. GPT-5.2 Pro, the company says, takes longer to generate answers but is its “smartest and most trustworthy” model for generating accurate answers in complex domains like programming.
For the many developers that are now developing agents, OpenAI says GPT-5.2 with reasoning is its strongest offering yet, bringing “significant improvements across general intelligence, long-context understanding, agentic tool-calling, and vision.”
OpenAI reportedly pushed to release GPT-5.2 before the end of the year so that it could counter the release of Google’s Gemini 3. The company released GPT-5 in August, heralding it as the next major leap forward in its AI research. GPT-5 was a “system” of models, using a “router” to direct the right queries to specialized models. It’s referring to GPT-5.2 as a “unified system that automatically chooses how to respond based on task complexity.”
The GPT-5.2 model’s increased capacity for processing and reasoning about multi-modal input (audio, video, images, text, etc.) is significant, because Google Gemini 3 does this very well.
For example, the new model was asked to analyze the features of an image of a circuit board and then identify and label all the small components. OpenAI says GPT-5.2 did this with far more detail and accuracy than its earlier GPT-5.1 model could. When reasoning is introduced, the model may be able to diagnose problems in mechanical systems by recognizing the visual signs.
All three variants of GPT-5.2 are available in ChatGPT today, starting with paid subscribers and available to developers through the API. Microsoft, a major investor in OpenAI, says it’s bringing GPT-5.2 to Microsoft 365 Copilot and Copilot Studio users worldwide today.
