OpenAI released GPT-5.2 on 11 December 2025 in three versions: Instant, Thinking and Pro. The headline number of the release was GDPval, a benchmark of real work tasks. In blind comparisons, experts judged the model’s output against a professional’s work, and GPT-5.2 Thinking came out no worse than the human in 70.9% of comparisons, Pro in 74.1%. GPT-5 had scored 38.8%. By September 2026 GPT-5.2 had already been removed from ChatGPT, its successors had climbed to almost 85% on GDPval, and OpenAI had reached the GPT-6 family.
What GDPval is and what the numbers mean
OpenAI published GDPval in September 2025. It contains 1,320 tasks from 44 occupations, from lawyers and engineers to nurses and video editors. The tasks were written by practitioners with an average of 14 years’ experience, and the output has to be a real piece of work: a spreadsheet, a presentation, a document. Experts from the same occupation compare it with a human’s work and decide whether it is better, as good or worse.
So 70.9% does not mean “the model is useful three times out of four”. It is the share of comparisons in which its result was judged no worse than an expert’s. Counting only outright wins, GPT-5.2 Thinking scores 49.8%. OpenAI itself flags a limitation: each task is done in a single pass, without edits or clarifications. Real work is different: results usually get reworked over several rounds.
The jump from 38.8% to 70.9% is still a big one. OpenAI called GPT-5.2 its first model to perform at expert level. That is the company’s own claim, and it has not been independently verified.
What else improved
- Errors in answers. By OpenAI’s measurements, 30% fewer answers contained errors, but that is with search switched on. Without search the difference is about 5%: 12.0% of answers with errors against 12.7% for GPT-5.1. Hallucinations, as OpenAI itself explains, stem from the way models are trained: a model is taught to guess the next word, and evaluations reward a confident answer over an honest “I don’t know”.
- Long context. On the MRCR v2 test, where the model has to find the right answer among similar ones in a huge text, GPT-5.2 is almost flawless on short documents. At 128–256 thousand tokens with eight “needles” it scores 77% against 29.6% for GPT-5.1. That is a huge gain, but the model still gets roughly one case in four wrong.
- Agentic tasks. On τ²-bench Telecom, where a model with tools handles support requests for a telecoms operator, GPT-5.2 scored 98.7%. The “from 47% to 98%” figure that went round the news compares different modes: at maximum reasoning the gain was from 95.6% to 98.7%, without reasoning from 47.8% to 57.2%.
- Code and maths. SWE-bench Verified rose from 76.3% to 80%, and the model solved 100% of the AIME 2025 problems.
One agent instead of many
The idea that GPT-5.2 lets you replace a fragile system of several agents with a single agent with twenty-odd tools did appear in the release. But it was a customer quote, from Triple Whale, not OpenAI’s own position. Tellingly, seven months later, with GPT-5.6, OpenAI itself shipped a mode with several agents working in parallel.
What came next
- 5 March 2026: GPT-5.4, with 83% on GDPval, a context of up to a million tokens and computer use.
- 23 April: GPT-5.5, with 84.9% on GDPval, a whisker short of 85%. Whether the next model crossed that line, the classic test can no longer tell us.
- 12 June: GPT-5.2 was removed from ChatGPT, and old conversations were moved to GPT-5.5.
- 9 July: GPT-5.6 in three versions, Sol, Terra and Luna. With it OpenAI stopped publishing the classic GDPval and switched to GDPval-AA v2, Artificial Analysis’s version of it with an Elo rating.
- 3 September: GPT-6 Astra; 22 September: GPT-6 Sol and Luna.
- 29 September: GPT-6.1 Sol, close to Astra in agentic coding and computer use but noticeably cheaper; 26 October: GPT-6.1 Terra, replacing GPT-5.6 Terra.
For how prompting changes with the new flagship, see our piece on GPT-6 Astra and short skills.
Has AI changed the office job market
So far, less than the benchmarks promised. The independent Yale Budget Lab tracker (last updated in October 2026) finds no clear sign of AI in employment or unemployment by occupation. But the impact shows up in specific places: according to a Stanford study based on ADP data (revised in August 2026), employment of 22–25-year-olds in the occupations most exposed to AI is 19% lower than it would be had it kept pace with less exposed occupations. Experienced workers show no such gap. AI is not yet laying people off; it is cutting back on hiring newcomers.
The bottom line
GPT-5.2 was not a quiet release but a turning point: for the first time, an OpenAI model matched experts on real work tasks in most comparisons. Nine months later that level had become the norm, and the model itself had become history. Even so, benchmark figures deserve a careful read: they show what a model can do in its best mode, not how it behaves in an ordinary chat.
How ChatGPT’s research mode was rebuilt on this model: Deep Research on GPT-5.2.
