OpenAI unveils GPT-5.2, its most advanced model for professional work

OpenAI has introduced GPT-5.2, describing it as its most capable model so far for professional tasks and long-running AI agents. The company says the new model family creates documents, spreadsheets, presentations, and code more accurately and much faster than earlier versions, while also understanding images and very long texts more reliably. In internal GDPval tests, which simulate real work across 44 types of jobs, GPT-5.2 Thinking matched or beat human experts in around 71 percent of comparisons, at a fraction of the time and cost.

GPT-5.2 shows strong gains in software development. On the SWE-Bench Pro benchmark of real-world coding tasks, it sets a new high score and does even better on the Python-focused SWE-Bench Verified test. Early partners such as Warp, JetBrains, and other developer tool makers report that the model is more dependable for debugging, refactoring, and building complex interfaces from a single prompt.

The model also improves on factual accuracy, long-context reasoning, and tool use. OpenAI says answer errors on real ChatGPT questions are about 30 percent less common than with GPT-5.1. On its MRCRv2 long-context test, GPT-5.2 can pull together information spread across up to 256,000 tokens, helping with work like reviewing long contracts, research papers, or multi-file projects. On tool-calling tests such as Tau2-bench Telecom, it completes nearly all multi-step tasks correctly, which should make automated workflows like customer support and data retrieval more reliable.

GPT-5.2 is also OpenAI’s best vision model to date, cutting error rates on chart and interface understanding, and performing strongly on scientific and math benchmarks such as GPQA Diamond and FrontierMath. The company highlights early examples where researchers used GPT-5.2 to help explore open questions in mathematics under human supervision.

The model comes in three main variants in ChatGPT: Instant for quick everyday use, Thinking for deeper work, and Pro for the highest-stakes tasks. GPT-5.2 is rolling out first to paid ChatGPT plans and is available now in the API under several model names, with higher prices than GPT-5.1 but advertised better efficiency per task. OpenAI says safety protections have also been strengthened, especially around mental health and sensitive content, and that further improvements are planned.