AI Vision Index
Reports
What we learned from the tests, in plain language. Every number comes from the leaderboard, and each report says when we ran the tests.
We tested 36 AI vision models on private testsAI Vision Index: 36 models from 18 makers. Who sees best, what it costs, and where every model still gets it wrong.Read the report
The best AI for reading gauges, labels and fine print is GPT-6 Astra, by a clear marginGPT-6 Astra scores 88 out of 100 on our eight reading tests, six points clear of the next model. On the smallest print it gets 98; Claude Opus 5.5 gets 44.
Can AI count past 20? Not in a cluttered picture.Up to 10 objects, the best models count right nine times in ten or more. With 21 to 30 in a busy scene, the best gets one in five. Neat stacks and one-at-a-time video are a different story.
Check an invoice: yes. Count a crowd: not yet. Which picture jobs AI can do todayOn four of our 29 tests, nine or more of the 24 models tested on our larger sample score 90 out of 100 or more, and the cheapest of them cost 20 to 74 cents per 1,000 answers. On five tests, none of them reaches 80.
The most expensive AI model we tested is not the bestClaude Fable 5.1 costs about $88 per 1,000 answers and ranks ninth of 36. Gemini 4 Argon costs $6.52 and scores 7.6 points higher.
The top AI vision model can barely tell an AI-made picture from a real photoGemini 4 Argon gets 81% right. GPT-6 Astra, first overall, gets 57.5% on a test where a coin flip gets 50%.
Video is where the best AI models pull away from the restOn our two video tests, five models score 89 or more and 15 of 34 score under 25. For the leaders our video tests are easier than the still pictures. For nearly everyone else they are much harder.
AI has mostly stopped seeing things that aren't there. Optical illusions still fool a third of the models.On questions about absent objects, 32 of 36 models score 85 or more out of 100. On optical illusions, 11 of 36 score 3 or less.
The best cheap AI vision model: 61 out of 100 for 74 cents per 1,000 answersFour models cost 20 to 74 cents per 1,000 answers. Their scores run from 12 to 61. In the cheap tier, which one you pick matters more than the price.
Small AI models that run on your own PC score a third of the best cloud modelFive small open models ran on one ordinary PC. The best scored 29.5 out of 100, against 83.7 for GPT-6 Astra. Google's cheapest cloud model, at 34 cents per 1,000 answers, beat all five. Bigger open models do better, but need a bigger machine.