What is GPT-5.2, and what changes for people who use ChatGPT for work?

Yesterday, Thursday, December 11, 2025, OpenAI released GPT-5.2, a new series of models for ChatGPT and for developers. OpenAI presents it as its most capable model yet for professional knowledge work: building spreadsheets and slide decks, writing code, reading images and long documents, using tools and carrying a complex project through several steps.

The release comes less than a month after GPT-5.1, which focused on making ChatGPT warmer and more pleasant to talk to. GPT-5.2 keeps that tone but puts the weight on accuracy and getting work done. Fidji Simo, OpenAI's CEO of Applications, told reporters the model had been in development for many months.

What OpenAI released

GPT-5.2 comes in three versions that share the same name but trade speed for depth:

In ChatGPT, the rollout began yesterday with paid plans (Plus, Pro, Go, Business and Enterprise), gradually, so some people will see it later than others. GPT-5.1 stays in the menu of older models for three months before it is retired. Creating spreadsheets and presentations inside ChatGPT is limited to the Plus, Pro, Business and Enterprise plans, using the Thinking or Pro version, and complex files can take many minutes.

For developers, all three are available now. GPT-5.2 costs $1.75 per million input tokens and $14 per million output tokens, more than GPT-5.1's $1.25 and $10. A token is a small chunk of text, roughly part of a word. OpenAI argues that the new model needs fewer tokens to reach a given quality, so a finished task can still cost less. The Thinking and Pro versions also gain a new top setting for reasoning effort, called xhigh, for jobs where quality matters most.

What the test scores mean

OpenAI backs the release with benchmark results, all reported by the company:

The previous model is GPT-5.1 Thinking unless noted.

GDPval is the headline. It gives the model clearly defined assignments drawn from 44 occupations, such as a sales presentation, an accounting spreadsheet or a schedule for an urgent care clinic. Expert judges then compare the model's work with work by experienced professionals. GPT-5.2 Thinking won or tied 70.9% of those comparisons, which OpenAI says makes it the first of its models to perform at or above expert level on this test. OpenAI adds that it produced these outputs more than 11 times faster than the professionals and at under 1% of the cost, while noting that human oversight still matters.

Other gains are practical. On anonymized ChatGPT questions, answers containing an error were about 30% less common than with GPT-5.1 Thinking. On a long-document test where the model must find and connect details scattered across hundreds of thousands of tokens, it is far more accurate than its predecessor. OpenAI also says it roughly halves error rates when reading charts and software screens, and that it uses tools more reliably across long conversations.

How it works in practice

OpenAI does not explain the model's design, but the three versions show the main idea. Instant answers right away. Thinking is a reasoning model: before replying, it works through the problem privately, step by step, which takes more time but catches more mistakes. Pro thinks longer still.

Here is the kind of request GPT-5.2 Thinking is built for, adapted from the sort of task OpenAI shows:

Build a spreadsheet to plan staffing for a 40-person company next year.
Include a sheet per department (engineering, sales, marketing, support),
monthly hiring, an expected 10% yearly turnover, and the effect on the
budget. Add a summary sheet with totals and one chart. List any
assumptions you made in a separate tab.

The last line matters: asking the model to state its assumptions makes its work easier to check.

Why it matters

The release sits in a tight race. The Verge reports that Sam Altman sent an internal "code red" memo the week before, telling staff to focus on improving ChatGPT and to delay other projects such as advertising, as Google's Gemini 3 gained ground. GPT-5.2 is OpenAI's first major answer.

The direction is clear: less chatbot, more coworker. OpenAI quotes companies such as Notion, Box, Shopify, Zoom and Databricks praising its handling of long, multi-step work, and it says a version tuned for its Codex coding agent will follow in the coming weeks.

OpenAI also reports progress on safety. Compared with GPT-5.1, both Instant and Thinking handle sensitive conversations better, such as signs of self-harm, emotional distress or an unhealthy reliance on the chatbot. The company is starting to roll out a system that estimates whether a user is under 18 so that teen protections apply automatically. And it is frank about limits: like every model, GPT-5.2 Thinking still makes mistakes, and critical answers should be double-checked.

Which GPT-5.2 version should you use, Instant, Thinking or Pro?

Use Instant for most everyday requests: quick facts, explanations, emails, translations and step-by-step how-tos. It is fast, and early testers noticed clearer explanations with the key point up front.

Switch to Thinking when the task has several parts or a cost to being wrong: analyzing a long document, writing or fixing code, working through numbers, or building a spreadsheet or presentation. It is slower, so give it a clear goal and the format you want.

Save Pro for the hardest questions, where you are happy to wait for a more careful answer, such as a technical review or a complex research question. Through the API it costs far more: $21 per million input tokens and $168 per million output tokens.

Whichever you choose, the habits from earlier models still apply: say what the result is for, ask the model to show its assumptions, and verify anything important before you use it.