50.6%
#1
ClockBench accuracy of GPT-5.4 High in March 2026, the best of the top 15; humans score about 90%
What the chart shows
ClockBench scales up clock reading to 180 clock designs and 720 questions. In March 2026 GPT-5.4 High led with 50.6%, ahead of Qwen 3-VL 235B Instruct (39.4%) and Gemini 3.1 Pro (32.2%). Humans read correctly formatted clocks about 90% of the time.
Ten of the 15 models scored 15% or less, including Claude Opus 4.6 at 8.9%. The report adds that when models got the time wrong, their median error was one to three hours, against three minutes for humans.
Key numbers
| Measure | Value |
|---|---|
| GPT-5.4 High | 50.6% |
| Qwen 3-VL 235B Instruct | 39.4% |
| Gemini 3.1 Pro | 32.2% |
| Gemini 3 Pro | 28.9% |
| Claude Opus 4.6 | 8.9% |
| Human baseline (report) | 90.1–90.7% |
How to read it
- Top 15 models on the ClockBench leaderboard (clockbench.ai), March 2026.
- The human baseline is not in the data file; the report’s text gives 90.1% and its chart 90.7%, so we say about 90% and draw the line at 90%.
- Qwen 3-VL 235B scored 39.4% in its instruct version and 14.4% in its thinking version.
Source: ClockBench Leaderboard, 2026, via the AI Index Report 2026 (Stanford HAI), figure 2.4.8: ClockBench: accuracy. Chart and text by AI Adoption.
Numbers checked against the published data on 20 October 2026.
Download the data (CSV) · Cite this chart
Numbers checked against the published data on 20 October 2026.
Download the data (CSV) · Cite this chart
Discussion