93%
▲ 57 pts since 2023
best mean accuracy on GPQA Diamond in 2025 (gemini-3-pro-preview), against 81.2% for expert human validators
20232025

What the chart shows

GPQA Diamond is a set of hard, text-only science questions written so that the answers cannot easily be looked up. In Epoch AI’s evaluations the best mean accuracy rose from 36% in 2023 (GPT-4) to 77% in 2024 (o1 at high reasoning effort) and 93% in 2025 (gemini-3-pro-preview).

Expert human validators score 81.2%, so the best model of 2025 was 11.8 points above them. The report says OpenAI’s o3 already passed the expert baseline in late 2024 with 87.7%; that result is not in this data.

Key numbers

MeasureValue
2025: gemini-3-pro-preview93%
2024: o1-2024-12-17 (high)77%
2023: GPT-4 (Mar 2023)36%
Expert human validators (report)81.2%

How to read it

  • Each bar is the highest mean accuracy that year in Epoch AI’s own runs, which can differ from developers’ reported scores.
  • The 81.2% expert baseline comes from the report’s chart; it is not in the data file.
  • The report text cites o3 at 87.7% in late 2024; Epoch’s data gives 77% for 2024, which we use.
Source: Epoch AI, 2026, via the AI Index Report 2026 (Stanford HAI), figure 2.4.2: GPQA on the diamond set: mean accuracy. Chart and text by AI Adoption.
Numbers checked against the published data on 20 October 2026.
Download the data (CSV) · Cite this chart