“Can AI do this job?” depends on the job. We sorted our 29 tests by two things: the best score any of the 24 models reached, and how many models score 90 or more out of 100. We use 90 as a rough bar for “you could rely on it, with spot checks”. Your own bar may be higher.
This page covers the 24 models tested on our larger sample; the leaderboard has all 36.
4 or more models at 90+No model reaches 80
The bar is the best score any model reached on the test (out of 100). The second figure is how many models score 90 or more.
Ready, with plenty of choice
On these four tests, nine or more models score 90 or more, so you can pick on price:
| Job | Models at 90 or more | Cheapest of them |
|---|---|---|
| Absent objects | 15 | Perceptron Mk1.5, $0.20 |
| Saying “I can’t tell” | 13 | GPT-6 Luna, $0.74 |
| Which bay can I park in? | 11 | GPT-6 Luna, $0.74 |
| Check the invoice | 10 | GPT-6 Luna, $0.74 |
That is less than a tenth of a cent per answer. For “is there a … in the picture?”, a small model running on your own PC, Qwen3-VL-8B, also scores 90.
Ready, but only with a top model
On these, three to seven models score 90 or more and the rest fall far behind, so the choice of model matters. The cheapest model at 90 or more is Gemini 4 Argon or GPT-6.1 Sol, at about $7 to $8 per 1,000 answers:
| Job | Models at 90 or more | Cheapest of them |
|---|---|---|
| Work out a missing measurement on a drawing | 7 | Gemini 4 Argon, $6.52 |
| Follow a pipe through a diagram | 5 | Gemini 4 Argon, $6.52 |
| Read a meter or dial gauge | 5 | Gemini 4 Argon, $6.52 |
| Compare a shelf with its plan | 5 | Gemini 4 Argon, $6.52 |
| Read an analogue clock to the minute | 5 | Gemini 4 Argon, $6.52 |
| Count items on a video conveyor | 4 | GPT-6.1 Sol, $7.94 |
| Follow a hidden object in a video | 4 | GPT-6.1 Sol, $7.94 |
| Follow one cable through a tangle | 3 | GPT-6.1 Sol, $7.94 |
Claude Fable 5.2 is the only model at 90 or more on all eight, but at $53.10 per 1,000 answers it is never the cheapest. The median model scores far lower on most of these, and a cheap model is a gamble.
Not ready
On these five tests, no model reaches 80:
| Job | Best model | Best score | Median model |
|---|---|---|---|
| Test strips | Claude Fable 5.2 | 77.5 | 49.6 |
| Same or mirror image | GPT-6 Astra, Claude Fable 5.2 | 75.0 | 2.4 |
| Rendering glitch | Gemini 4 Argon | 73.3 | 32.9 |
| Real or AI-made | Gemini 4 Argon | 61.3 | 22.7 |
| Counting objects | GPT-6 Astra, Claude Fable 5.2 | 52.5 | 27.2 |
Counting objects in a busy scene is the clearest case: AI counts small numbers well and loses count past 20. Telling an AI-made picture from a real photo is another: only one model is clearly better than a coin flip.
What it means
Yes-or-no checks with a clear answer in the picture, such as “is it there?”, “can you tell?” or “does this invoice add up?”, are close to solved, and cheap. Jobs that need a fine visual judgement, such as faint lines, mirror images, small rendering errors or exact counts in clutter, are not, and a person should check them. In between, pay for a top model and test it on your own pictures.
How we tested
Each test asks one kind of question with one exact answer built into the picture, at difficulty levels from easy to hard. Answers are scored automatically and corrected for guessing, so 0 is a guess and 100 is always right. “Best score” is the highest any model reached on the test. Some models, including the most expensive ones, were run on an evenly spread sample of each test, so their single-test scores have wider margins of error. The pictures and questions are never published.
Tested 22 October to 10 November 2026 (the 24 models on this page). Each model’s dates are on the leaderboard.