Where does AI perform best today?

Competition mathematics is the use where the level is best established. In July 2025, a version of Gemini known as Deep Think solved 5 of the 6 problems of the International Mathematical Olympiad, in natural language and within the 4 h 30 of the exam; its papers, graded by the Olympiad markers with the pupils' criteria, scored 35 points out of 42, the score of a gold medal. The limit is research: on 68 open problems posed by the mathematician Paul Erdős, selected by Epoch AI, the best model had solved 3 in September 2026.

Code comes next. On SWE-bench, a benchmark that asks models to fix real bugs in open-source software written in Python, the best models solved 4.4% of problems in 2023 and 71.7% in 2024. On SWE-bench Verified, a selection of 500 of these problems, the 2026 edition of the Stanford AI Index reports a rise from 60% to nearly 100% in a year. The real work of developers is harder to measure: in the METR controlled trial, experienced open-source developers took 19% more time with the tools of early 2025, and the February 2026 update does not allow the effect of more recent tools to be quantified, as detailed in our analysis of productivity. Code is also the use where the duration of successful tasks grows fastest, as measured by METR (our analysis of the speed of progress).

These 2 uses have one thing in common: the result can be checked. A proof can be reread, a fix either passes the tests or does not.

Writing is the use where controlled trials measure the clearest gains, with models from 2023. Graduate professionals who had to write a press release, a memo or a delicate email took 40% less time with ChatGPT, and their texts were rated higher by professionals in the same occupation. In a customer service company, the assistant led to 14% more requests being resolved per hour, 34% among beginner agents, with little effect on the most experienced.

Office deliverables are measured by GDPval, an OpenAI benchmark: 220 tasks from 44 occupations (a presentation, a spreadsheet, a document), built by professionals with 14 years of experience on average, which take a human nearly 7 hours on average. Experts in the occupation compare, blind, the model's deliverable and that of an expert. In September 2025, 47.6% of the deliverables of Claude Opus 4.1 were judged as good or better; those of GPT-5 were judged better, not counting ties, in 39.0% of cases. OpenAI specifies that the tasks are described in advance and handled in a single pass, without interaction, whereas real work often requires reconstructing the context.

Information retrieval does not have one level, it has several. A model makes few mistakes when it summarises a text it is given (3 to 15% of erroneous summaries), more when it answers from memory, and many when it has to find the source of an article (more than 60% of incorrect answers). Our analysis of the reliability of answers details these measurements. We have found no public measurement on the analysis of long documents of several hundred pages.

Judgement outside the model's domain is the only case on this scale where AI lowered the result, on a task built to make it fail. In 2023, 758 Boston Consulting Group consultants handled tasks with or without GPT-4. On 18 tasks within the model's reach, those who used it finished 12.2% more, 25.1% faster. On a business case designed so that GPT-4 would go wrong, where the spreadsheet alone led to a wrong conclusion that details from the interviews corrected, they gave the right answer in 60% and 70% of cases depending on the group, against 84.5% without AI. The authors speak of a jagged frontier: the model fails on tasks that seem to be of the same difficulty as those it succeeds at. The measurement dates from GPT-4 and has not been repeated identically.

Long projects carried out alone are the least successful. The Remote Labor Index, published in October 2025 by Scale AI and the Center for AI Safety, gives agents 240 real freelance projects, which had required more than 6,000 hours of work. In the study, with deliverables produced mainly in June 2025, the best agent produced a result judged acceptable for 2.5% of projects. The Scale AI ranking, which rates agents on the same 230 unpublished projects, gives the best model 41% in November 2026, with no measurement date.

The frequency of a use and its measured performance do not follow the same order. In ChatGPT, from May 2024 to June 2025, requests for practical advice, information retrieval and writing make up about 77% of conversations, code 4.2% of messages, according to a study by OpenAI and researchers from Harvard and Duke published by the NBER. At work, writing makes up 40% of messages, and 2 thirds of writing requests concern a text that the user supplies. Information retrieval is among the most common uses and among those whose measured error rates vary the most.