People now use AI models to check whether a picture is real. Our tests suggest you should not rely on them, with one partial exception.
In the “Real or AI-made” test, a model sees one picture and says whether it is a real photograph or was generated by an image model. Half the pictures are real. A model that guesses is right half the time. Here is how often each of the leading models got it right:
| Model | Maker | Share right |
|---|---|---|
| Gemini 4 Argon | 80.7% | |
| Claude Fable 5.1 | Anthropic | 74.0% |
| Claude Opus 5.5 | Anthropic | 72.0% |
| Claude Sonnet 5.5 | Anthropic | 72.0% |
| Claude Fable 5.2 | Anthropic | 70.0% |
| GPT-6 Luna | OpenAI | 62.0% |
| GPT-6.1 Sol | OpenAI | 59.3% |
| GPT-6 Astra | OpenAI | 57.5% |
GPT-6 Astra and GPT-6.1 Sol, first and joint second on the index overall, are barely better than a coin flip here. Astra’s margin of error is wide enough to include 50%. Only Gemini 4 Argon is above 80%. Of the 24 models tested on our larger sample, nine are below 60%, and Mistral Small 2603 is at exactly 50%.
The other 12 models, tested on a smaller sample, are on the leaderboard.
Each bar is the share of pictures a model called correctly; the whisker is the 95% interval. Half is what a coin flip gives.
The cliff
The pictures come in difficulty levels. On the first three, the average model manages 70%, 76% and 71% right. On the last two it falls to 50% and 46%. That is a coin flip.
Thin grey lines: one per model. Level 1 is the easiest. The dashed line marks a coin flip at 50.
Gemini 4 Argon holds up best on the hardest levels, with 63% right on both. At Levels 1 to 3 it gets 100%, 87% and 90%. Claude Opus 5.5 goes 93%, 87%, 83%, then 47% and 50%. GPT-6 Astra starts lower and falls to 25% on the last level, which is worse than guessing. Claude Fable 5.2 gets 88% on each of the first three levels, then 62%, and also 25% on the last. Each level is a small sample, so do not read much into one number.
What it means
If a model says “this looks real” or “this looks AI-generated”, it is right about four times in five at best, and that is across all levels together. On the hardest levels, most models are guessing. The model that is best at nearly everything else, GPT-6 Astra, ranks 17th of these 24 models on this test.
How we tested
Each picture is either a real photograph or one made by an image generator, shown on its own. We paired each AI-made picture with a real photo of a similar subject, so a model cannot succeed by knowing what typical subjects look like. The question is always the same and the answer is exact. Scores here are the raw share of right answers, because the chance of a guess is simply 50%. On the leaderboard the same test is shown corrected for guessing: 0 for a coin flip, 100 for always right. On that scale Gemini 4 Argon scores 61, and GPT-6 Astra 15. The pictures are never published.
Tested between 22 October and 10 November 2026. All dates are on the leaderboard.