Skip to main content
Back to timeline
OpenAISource publication:

OpenAI says ChatGPT now handles about 8.4 million weekly hard-science prompts, as research use spreads from writing to derivation and checking

Synopsis

OpenAI reports that weekly messages on advanced science and math topics on ChatGPT rose from 5.7 million to nearly 8.4 million during 2025, about 47%, with roughly 1.3 million weekly users discussing such topics as of January 2026. The report sets those usage figures alongside specific cases in mathematics, physics, chemistry and biology, where models are used for literature synthesis, code debugging, derivations and experiment planning; the numbers come from OpenAI's internal sampling of anonymized conversations and measure usage scale, not independently assessed research output.

Source-provided article image: AI as a Scientific Collaborator January 2026
OpenAI · Page 1

Interpretation

Use of ChatGPT for advanced science and math topics expanded markedly within a year, and OpenAI separates this cohort as research-oriented users. The comparison the report gives is the start and end of the same year: about 5.7 million weekly messages in 2025 rising to nearly 8.4 million by year's end, roughly 47%. The comparison comes from OpenAI's internal analysis of a full random sample of anonymized ChatGPT conversations from January through December 2025, with nearly 1.3 million weekly users discussing these topics as of January 2026.

This cohort uses the system differently from typical users, concentrating on coding, data analysis, mathematical reasoning and literature review. Against typical ChatGPT users as baseline, the report says advanced science and math users send about 3.5 times more messages, send coding-related messages nearly 12 times as often, and average 9 informational-overview prompts per week versus 1.5. These ratios also come from OpenAI's internal analysis of the same anonymized conversations, and the report defines advanced prompts as research-oriented questions requiring graduate-level or research-level expertise.

In mathematics, the report attributes progress to test-time compute scaling and training that rewards verifiable outcomes, and gives checkable benchmark figures. The report says GPT-5.2 Thinking scored perfectly on AIME 2025 without external tools and solved 40.3% of FrontierMath problems in Tiers 1-3, while GPT-5.2 Pro scored 31% on the harder FrontierMath Tier 4. The report also states that FrontierMath is built to be accessible only to true experts and describes Tier 4 problems as mini research projects, so the two scores belong to problem sets of different difficulty.

In biology, OpenAI and RetroBioSciences used a protein-focused foundation model, GPT-4B Micro, to engineer OSKM reprogramming factors. The report says traditional protein engineering of super variants has taken years of trial and error, while this collaboration compressed iteration cycles from years to months. The model generated thousands of candidate sequences, filtered to preserve essential domains and diversity, then narrowed to a few hundred variants that RetroBio synthesized, delivered into human fibroblasts by lentiviral constructs and reprogrammed for roughly 10-14 days, ranking variants by cell-surface markers and sequencing; the report says about half of model-generated variants outperformed wild-type proteins for some targets.

Perspective

The report presents OpenAI's internal statistics on ChatGPT usage plus several collaboration cases, aimed at researchers and engineering teams watching how AI enters research workflows; the figures describe conversation volume and user counts, not independent measurements of research output or correctness.

The report says some AI contributions in mathematics still amount to stitching known methods together, and inventing entirely new kinds of mathematics is beyond current models; the ChemBench chemistry benchmark also shows models still failing some basic tasks and producing overconfident errors, requiring human oversight.