Reading is the most common job people give an AI that looks at pictures: the value on a gauge, a serial number on a label, a figure on a chart, the line on a rapid test. Our Reading category has eight such tests, each with one exact answer. On all of them together, one model is clearly ahead.
The ranking
GPT-6 Astra from OpenAI scores 88 out of 100 in Reading. Its 95% interval runs from 85 to 91. Three models follow close together: GPT-6.1 Sol scores 82 (80 to 84), Claude Fable 5.2 81 (78 to 85) and Gemini 4 Argon 80 (78 to 81). No other model’s interval reaches Astra’s, so this is not a tie. Claude Opus 5.5 follows with 75. Kimi K3.1 Pro from Moonshot (72) and Qwen3.8 Max from Alibaba (67) were tested on a smaller sample, so their intervals are wide (64 to 80 and 59 to 75). Kimi K3.1 Pro could be as good as Gemini 4 Argon, and Qwen3.8 Max could be as good as Claude Opus 5.5.
US labsChinese labsOpen models via APIRuns locallyOther labs
The figure at the end of each bar is the score, then the cost per 1,000 answers.
Where Astra pulls ahead
From here on, single tests compare only the 24 models tested on a larger sample: with few answers, one test score swings too much. On easy reading, the top models are close. Telling the time on an analogue clock: GPT-6.1 Sol 98, GPT-6 Astra 98, Gemini 4 Argon 96. The gap opens on small print, charts and coins:
| Test | GPT-6 Astra | Best other model | Median model |
|---|---|---|---|
| Eye chart (smaller and smaller print) | 98 | 88 (GPT-6.1 Sol) | 23.8 |
| Reading a chart on a screen | 91 | 80 (GPT-6.1 Sol) | 47.5 |
| Adding up coins | 80 | 74 (GPT-6.1 Sol) | 40.2 |
| Test strips (faint positive lines) | 75 | 77.5 (Claude Fable 5.2) | 49.6 |
The eye chart is the clearest case. As the print gets smaller, GPT-6 Astra keeps reading it, and almost every other model gives up or guesses. Claude Opus 5.5 scores 44 there and Claude Sonnet 5.5 32, though both read labels and meters well. If your pictures have small text, this one test matters more than the overall leaderboard. On the faint lines of rapid tests, nobody is far ahead: Claude Fable 5.2 scores 77.5, and five models score between 70 and 75.
Cheaper choices
GPT-6 Astra costs $38.32 per 1,000 answers. If that is too much:
- GPT-6.1 Sol scores 82 in Reading for $7.94, about a fifth of the price.
- Gemini 4 Argon scores 80 for $6.52, and scores 89 on reading a code off a label in a scene, one point behind the best there, Claude Fable 5.2 (90).
- Claude Haiku 5.5 scores 62.5 for $1.01, and GPT-6 Luna 59 for $0.74. Fine for clear, large text; weak on small print.
Thin grey lines: one per model. Level 1 is the easiest.
What it means
For reading jobs, pick GPT-6 Astra if every answer must be right and the text can be small. Pick GPT-6.1 Sol or Gemini 4 Argon if your text is clear and the price matters. Test the model on your own pictures either way: on the faint lines of a rapid test, no model scores above 80.
How we tested
Each reading picture is built by script or in 3D with a known answer: a time to the minute, a value, a code. Pictures get harder in steps (smaller, blurrier, more glare, more clutter). Answers are checked automatically against the exact key and scores are corrected for guessing, so 0 is a guess and 100 is always right. The Reading score is the average of the eight test scores. The pictures and questions are never published.
Tested 22 October to 10 November 2026. Each model’s dates are on the leaderboard.