OpenAI launches GPT 5.2 with advanced research capabilities

OpenAI has unveiled GPT 5.2, a new line of frontier models aimed at professional work, long-running tasks, and complex reasoning. The release spans three versions, Instant, Thinking, and Pro, rolling out first to paid ChatGPT users and already available through the API. The company positions  GPT 5.2 as a major leap in capability, with stronger performance in coding, analysis, long-context workflows, and scientific problem-solving.

A stronger grip on knowledge work

OpenAI says GPT 5.2 Thinking now matches or surpasses human experts in most well-specified tasks measured through GDPval, a benchmark covering 44 professional occupations. Judges rated GPT 5.2 Thinking as beating or tying industry specialists in 70.9 percent of comparisons. Early outputs shared by evaluators suggested clearer structure, tighter formatting, and faster delivery, especially for presentations and spreadsheet models.

Internal testing also shows gains in financial modeling. On junior investment-banking tasks, such as building three-statement models or leveraged buyout models, average scores jumped nearly 10 percentage points over GPT 5.1.

Coding accuracy moves up

GPT 5.2 shows marked improvements in software engineering benchmarks. The model reached 55.6 percent on SWE-Bench Pro, which tests real repositories and multiple languages. On SWE-bench Verified, it reached 80 percent accuracy. According to early partner feedback from Windsurf, JetBrains, and Warp, engineers saw more reliable debugging, cleaner feature implementations, and stronger performance in interface-heavy front-end tasks.

The company shared demos showing GPT 5.2 generating single-file apps with 3D elements, mini-games, or animated simulations from a single prompt.

In de-identified ChatGPT queries, GPT 5.2 Thinking produced 30 percent fewer incorrect answers than GPT 5.1. OpenAI frames this as noticeable for research, analysis, and writing support, though it continues to warn that the model is far from perfect.

Long-context handling hits new highs

One of the biggest changes comes from how the model processes long documents. GPT 5.2 reached near-perfect scores on the 256k-token version of OpenAI’s MRCRv2 test, which requires tracking information across very long texts. This boost is meant to support heavy tasks such as reviewing large reports, analysing contracts, and working across multi-file projects without losing coherence.

For jobs extending beyond the context window, GPT 5.2 Thinking can also tap into the new /compact endpoint, which expands effective memory for tool-driven workflows.

Better at charts, diagrams, and interfaces

Vision performance also improves. GPT 5.2 Thinking nearly doubles accuracy in software interface understanding and raises scores on chart interpretation. It shows a clearer sense of spatial layout, helping identify components in hardware images or elements in dashboards.

Gains in science and math

GPT 5.2 Pro and GPT 5.2 Thinking rank among the strongest models yet on complex science and math benchmarks. GPT 5.2 Pro achieved 93.2 percent on GPQA Diamond, while GPT 5.2 Thinking set a high of 40.3 percent on FrontierMath Tier 13. Both models also advanced abstract-reasoning scores on ARC-AGI.

Pricing and availability

GPT52 is available immediately in the API under GPT 5.2 and gpt52chatlatest, with Pro accessible through gpt52pro. Pricing is higher than the GPT 5.1 series, starting at USD 1.75 per million input tokens and USD 14 per million output tokens. ChatGPT subscription prices remain unchanged.

The company will keep GPT 5.1, GPT 5, and GPT 4.1 active in the API for now, with deprecations announced well in advance. A version optimised for Codex-style coding tasks is expected soon.

Built with Microsoft and NVIDIA

OpenAI says the model was trained on Azure infrastructure backed by NVIDIA’s H100, H200, and GB200NVL72 GPUs, enabling faster scaling and shorter training cycles.

GPT 5.2 marks another step in the company’s march toward models that can handle sustained, end-to-end workflows. As OpenAI puts it, development continues, with work underway on reliability issues and safety improvements for younger users and sensitive conversations.