ClockBench logo
OverviewLeaderboardDatasetContact

Visual Reasoning Benchmark

Clock Bench

ClockBench evaluates whether models can read analog clocks - a task that is trivial for humans, but many AI models struggle with.

Clock Faces
36
Clocks
180
Questions
720
Top Human Accuracy
95%
Top Model Accuracy
83.3%
Read the Paper

Leaderboard

RankModelAccuracyLab
Top Human Performance95%
Average Human Performance90.7%
1Gemini 4 Argon High83.3%Google
2Claude Opus 5.5 Max77.2%Anthropic
3GPT-6 Astra Max75%OpenAI
4GPT-6.1 Sol Max74.4%OpenAI
5GPT-5.6 Sol Max68.3%OpenAI
6Gemini 3.8 Flash66.7%Google
7Claude Opus 5 Max65.5%Anthropic
8Claude Fable 5.1 Max59.4%Anthropic
9Gemini 3.7 Flash53.3%Google
10GPT-5.4 High50.6%OpenAI
11GPT-5.5 High46.1%OpenAI
12Muse Spark 1.342.2%Meta
13Qwen 3-VL 235B Instruct39.4%Alibaba
14Claude Fable 537.2%Anthropic
15Gemini 3.5 Flash34.4%Google
16Gemini 3.1 Pro32.2%Google
17Gemini 3 Pro31.1%Google
18Grok 4.521.7%xAI
19Grok 4.620%xAI
20Gemini 2.5 Pro18.9%Google
21GPT-5.2 High15.6%OpenAI
22Gemini Robotics ER 1.515%Google
23Claude Opus 4.715%Anthropic
24Qwen 3-VL 235B Thinking14.4%Alibaba
25o3 Pro14.4%OpenAI
26o3 High12.2%OpenAI
27Gemini 2.5 Flash11.1%Google
28GPT-5 High11.1%OpenAI
29GPT-5 Pro11.1%OpenAI
30Mistral Medium 3.110.6%Mistral
31Claude Opus 4.610%Anthropic
32GPT-5 Mini8.9%OpenAI
33Claude Opus 4.18.9%Anthropic
34Claude Sonnet 4.57.8%Anthropic
35Claude Sonnet 46.7%Anthropic
36GTP-4o6.7%OpenAI
37Qwen 2.5-VL 72B6.1%Alibaba
38GTP-5 Nano3.9%OpenAI
39Grok 4 Fast3.3%xAI
ClockBench AI Benchmark
Top Human Performance
95.0%
Average Human Performance
90.7%
Gemini 4 Argon High
83.3%
Claude Opus 5.5 Max
77.2%
GPT-6 Astra Max
75.0%
GPT-6.1 Sol Max
74.4%
GPT-5.6 Sol Max
68.3%
Gemini 3.8 Flash
66.7%
Claude Opus 5 Max
65.5%
Claude Fable 5.1 Max
59.4%
Gemini 3.7 Flash
53.3%
GPT-5.4 High
50.6%
GPT-5.5 High
46.1%
Muse Spark 1.3
42.2%
Qwen 3-VL 235B Instruct
39.4%
Claude Fable 5
37.2%
Gemini 3.5 Flash
34.4%
Gemini 3.1 Pro
32.2%
Gemini 3 Pro
31.1%
Grok 4.5
21.7%
Grok 4.6
20.0%
Gemini 2.5 Pro
18.9%
GPT-5.2 High
15.6%
Gemini Robotics ER 1.5
15.0%
Claude Opus 4.7
15.0%
Qwen 3-VL 235B Thinking
14.4%
o3 Pro
14.4%
o3 High
12.2%
Gemini 2.5 Flash
11.1%
GPT-5 High
11.1%
GPT-5 Pro
11.1%
Mistral Medium 3.1
10.6%
Claude Opus 4.6
10.0%
GPT-5 Mini
8.9%
Claude Opus 4.1
8.9%
Claude Sonnet 4.5
7.8%
Claude Sonnet 4
6.7%
GTP-4o
6.7%
Qwen 2.5-VL 72B
6.1%
GTP-5 Nano
3.9%
Grok 4 Fast
3.3%

Results Summary

Despite frontier models showing strong reasoning skills, mathematical ability, and visual understanding on multiple benchmarks, it took them a while to improve at reading analog clocks.

One hypothesis might be that this task sets a high bar for doing reasoning within the visual space, as opposed to text space.

However, LLMs still were able to close the gap with humans relatively quickly without drastic changes in their architecture.

Dataset

Sample Clocks

Few examples of clocks that we used in the benchmark.

Sample clocks from ClockBench

Questions

  1. Reading Time
    Models are asked to determine whether a given clock shows a valid time. If valid, they should report the hours, minutes, seconds, date, month, and day of the week (based on what is present), in a structured JSON format.
  2. Adding or Subtracting Time
    Models are asked to add or subtract varying amounts of time.
  3. Rotating Hands
    Models are asked to rotate one of the hands (hour, minute, or second) by a specified angle, clockwise or counterclockwise.
  4. Shifting Time Zone
    Models are asked to assume they are in New York during summer and report the corresponding time in various locations worldwide.

Try Yourself

Interesting in trying out ClockBench?
A small public dataset and sample evaluation code is available to everyone.

Public Dataset
Alek Safar
LinkedInX.com

Please reach out to [email protected] with ideas, suggestions, questions or any other inquiries.

Read the Paper