AI Vision Index is an independent leaderboard for AI models that look at pictures and video. We build the tests ourselves, mostly by script, so we know the right answer to every question exactly. The tests are private. We never publish the pictures or the questions, so models cannot have learned them from the web. Every result carries the date it was tested.
The index covers 36 models from 18 makers on 29 tests in six categories. All were tested between 22 October and 10 November 2026.
The short version
- GPT-6 Astra from OpenAI is first, with 83.7 out of 100. GPT-6.1 Sol (81.3) and Claude Fable 5.2 from Anthropic (80.7) share second place: their margins overlap. Gemini 4 Argon from Google (77.2) and Claude Opus 5.5 from Anthropic (76.5) follow. Only these five score above 75.
- Next comes a close group. Claude Sonnet 5.5 from Anthropic scores 71.0. Kimi K3.1 Pro from Moonshot and Qwen3.8 Max from Alibaba, the best Chinese models, are level at 70.6, just behind it. Both were tested on a smaller sample, so their margins are wider: plus or minus 4.2 for Kimi K3.1 Pro (66.6 to 75.0) and 4.5 for Qwen3.8 Max (66.1 to 75.2). Both ranges take in the scores of Claude Sonnet 5.5 and Claude Fable 5.1 (69.6), so all four are within each other’s margins. Together they share sixth to ninth place.
- The middle is wide and the bottom is low. Sixteen models score above 50. The median is 45.6. The lowest, LFM2.5-VL 1.6B, scores 12.4.
- No model is best at everything. Among the 24 models tested on our larger sample, GPT-6 Astra has the top score on 13 of the 29 tests and Claude Fable 5.2 on 10; three of those they share. Gemini 4 Argon leads the Spotting changes category.
- Price does not predict quality. The most expensive model we tested is ninth of the 36.
| Rank | Model | Score · 95% interval |
Cost per 1,000 answers | Tested |
|---|---|---|---|---|
| 1 | GPT-6 AstramediumOpenAI · API | 83.7±2.0 95% interval 81.7 to 85.7 | $38.32 | |
| 2–3 (tied) | GPT-6.1 SolmediumOpenAI · API | 81.3±1.3 95% interval 80.0 to 82.5 | $7.94 | |
| 2–3 (tied) | Claude Fable 5.2defaultAnthropic · API | 80.7±2.1 95% interval 78.6 to 82.8 | $53.10 | |
| 4–5 (tied) | Gemini 4 ArgondefaultGoogle · API | 77.2±1.3 95% interval 75.9 to 78.5 | $6.52 | |
| 4–5 (tied) | Claude Opus 5.5defaultAnthropic · API | 76.5±1.8 95% interval 74.8 to 78.3 | $26.16 | |
| 6–9 (tied) | Claude Sonnet 5.5defaultAnthropic · API | 71.0±2.2 95% interval 68.8 to 73.1 | $11.60 | |
| 6–9 (tied) | Kimi K3.1 ProdefaultMoonshot · API | 70.6±4.2 95% interval 66.6 to 75.0 | $71.79 | |
| 6–9 (tied) | Qwen3.8 MaxdefaultAlibaba · API | 70.6±4.5 95% interval 66.1 to 75.2 | $27.40 | |
| 6–9 (tied) | Claude Fable 5.1defaultAnthropic · API | 69.6±2.1 95% interval 67.5 to 71.7 | $87.95 | |
| 10–18 (tied) | Claude Haiku 5.5defaultAnthropic · API | 63.5±1.5 95% interval 62.1 to 65.0 | $1.01 | |
| 10–18 (tied) | Gemini 3.8 FlashmediumGoogle · API | 61.9±1.8 95% interval 60.2 to 63.7 | $12.80 | |
| 10–18 (tied) | GPT-6 LunamediumOpenAI · API | 60.7±1.5 95% interval 59.3 to 62.2 | $0.74 | |
| 10–18 (tied) | Qwen3.8-27BdefaultAlibaba · API | 59.0±8.1 95% interval 51.8 to 68.0 | $7.75 | |
| 10–18 (tied) | Qwen3.8 FlashdefaultAlibaba · API | 56.4±1.4 95% interval 55.0 to 57.8 | $3.04 | |
| 10–18 (tied) | Grok 4.7defaultxAI · API | 54.9±8.0 95% interval 48.1 to 64.1 | $27.40 | |
| 10–18 (tied) | Mistral Large 4defaultMistral · API | 52.7±7.5 95% interval 46.7 to 61.7 | $9.59 | |
| 10–18 (tied) | Muse Spark (open)defaultMeta · API | 47.6±8.5 95% interval 40.3 to 57.3 | $19.12 | |
| 10–18 (tied) | GLM-5V-TurbodefaultZhipu · API | 46.8±2.4 95% interval 44.6 to 49.5 | $12.50 | |
| 19–29 (tied) | DeepSeek V4.1 FlashdefaultDeepSeek · API | 44.4±1.3 95% interval 43.2 to 45.8 | $1.43 | |
| 19–29 (tied) | GLM-5.3 FlashdefaultZhipu · API | 43.2±1.5 95% interval 41.7 to 44.7 | $1.40 | |
| 19–29 (tied) | Qwen3.6-35B-A3BdefaultAlibaba · API | 43.0±7.8 95% interval 36.9 to 52.4 | $3.77 | |
| 19–29 (tied) | Qwen3-VL-235BdefaultAlibaba · API | 41.2±7.6 95% interval 35.0 to 50.2 | $7.96 | |
| 19–29 (tied) | MiniMax M3defaultMiniMax · API | 40.9±2.5 95% interval 38.7 to 43.6 | $3.53 | |
| 19–29 (tied) | Step-5 (preview)defaultStepFun · API | 40.7±8.6 95% interval 33.2 to 50.3 | $12.50 | |
| 19–29 (tied) | MiMo v2.6 ProdefaultXiaomi · API | 40.3±2.0 95% interval 38.6 to 42.6 | $4.60 | |
| 19–29 (tied) | Gemini 3.5 Flash-LitedefaultGoogle · API | 40.3±1.5 95% interval 38.7 to 41.8 | $0.34 | |
| 19–29 (tied) | Gemma 4 31BdefaultGoogle · API | 38.8±6.6 95% interval 33.4 to 46.6 | $1.94 | |
| 19–29 (tied) | Seed 2.1 TurbodefaultByteDance · API | 38.3±7.1 95% interval 32.3 to 46.5 | $10.85 | |
| 19–29 (tied) | Hunyuan HY4defaultTencent · API | 37.9±7.7 95% interval 31.2 to 46.5 | $5.71 | |
| 30–31 (tied) | Qwen3.5-9Bno thinkingAlibaba · runs locally | 29.5±1.5 95% interval 28.0 to 31.1 | $0 (local) | |
| 30–31 (tied) | Perceptron Mk1.5defaultPerceptron · API | 28.8±1.5 95% interval 27.4 to 30.4 | $0.20 | |
| 32 | Qwen3-VL-8BdefaultAlibaba · runs locally | 26.0±1.4 95% interval 24.8 to 27.6 | $0 (local) | |
| 33 | MiniCPM-V 4.5defaultOpenBMB · runs locally | 17.1±1.4 95% interval 15.9 to 18.7 | $0 (local) | |
| 34–36 (tied) | Gemma 4 E4BdefaultGoogle · runs locally | 13.7±1.3 95% interval 12.5 to 15.2 | $0 (local) | |
| 34–36 (tied) | Mistral Small 2603defaultMistral · API | 12.5±1.4 95% interval 11.5 to 14.2 | $0.26 | |
| 34–36 (tied) | LFM2.5-VL 1.6BdefaultLiquid · runs locally | 12.4±1.4 95% interval 11.4 to 14.2 | $0 (local) |
The overall score is the average of the six category scores. Ranks shown as “2–3” are tied: those scores are too close to tell apart.
The six categories
Each test belongs to one category. A model’s category score is the average of its test scores. Its overall score is the average of the six category scores, so no single category dominates.
| Category | What it covers | Best model | Score | Average of all 36 |
|---|---|---|---|---|
| Reading | Labels, meters, charts, codes, traffic lights, clocks | GPT-6 Astra | 88 | 50 |
| Counting | Objects in a scene, stacked goods, items on a video conveyor | GPT-6 Astra | 78 | 33 |
| Space, 3D and motion | Depth, direction, mirror images, toppling stacks, following a hidden object | Claude Fable 5.2 | 86 | 44 |
| Spotting changes and errors | What changed, what is wrong, what does not match | Gemini 4 Argon | 89 | 56 |
| Diagrams | Following cables and pipes, working out a missing measurement | GPT-6 Astra | 98 | 54 |
| Not being fooled | Absent objects, illusions, AI-made pictures, saying “I can’t tell” | Claude Sonnet 5.5 | 78 | 59 |
GPT-6 Astra leads three of the six categories. On Space, 3D and motion, Claude Fable 5.2 (86) and GPT-6 Astra (84) are within each other’s margins. Not being fooled is the closest race: Gemma 4 31B from Google, tested on a smaller sample, also scores 78 there.
Counting is the weakest category for 26 of the 36 models. Diagrams is the easiest at the top: the two best models score 98 and 96 there.
What surprised us
The priciest model is not the best. Claude Fable 5.1 costs about $88 per 1,000 answers and ranks ninth of the 36, level with the cheaper Claude Sonnet 5.5 and behind Gemini 4 Argon, which costs $6.52 for the same work. Claude Fable 5.2 costs $53.10 and shares second place. Read the finding.
The leader cannot tell a real photo from an AI-made one. GPT-6 Astra gets 57.5% right on a test where guessing gets 50%. Of the 24 models tested on our larger sample, only Gemini 4 Argon is clearly good, at 81%. Read the finding.
AI has stopped inventing objects that are not there, but illusions still fool it. On questions about absent objects, 32 of 36 models score 85 or more. On optical illusions, 11 of 36 score 3 or less. Read the finding.
Video is where the leaders pull away. Across our two video tests, the top two models score 99 and 98. Fifteen of the 34 models that took both score under 25. Read the finding.
Chinese models reach the top ten, but most are far behind. Kimi K3.1 Pro and Qwen3.8 Max are level at 70.6, in the group that shares sixth to ninth place. Qwen3.8 Max costs $27.40 per 1,000 answers. Kimi K3.1 Pro costs $71.79, the second highest price of all. The other nine models from Chinese labs score between 37.9 and 56.4. The median of the 13 models from US labs is 69.6; for the 11 from Chinese labs it is 43.2.
The models you can run on your own PC are far behind. The best of the five we ran locally scores 29.5. Read the finding.
The cheap end hides a five-fold spread. Under a dollar per 1,000 answers, scores run from 12.5 to 60.7. Read the finding.
Cost against score
We report the cost of 1,000 answers at standard prices, where an answer is one question about one picture or clip. It is the list price of the tokens a model used, including its own thinking. That is the same for every model, whether we called it through an API or through its maker’s command-line tool.
The best deals sit top-left on the chart. GPT-6.1 Sol scores 81.3 for $7.94. Gemini 4 Argon scores 77.2 for $6.52. GPT-6 Luna scores 60.7 for $0.74. Qwen3.8 Max costs about the same as Claude Opus 5.5 ($27.40 against $26.16) but scores 70.6 against 76.5. Models running on our own PC cost nothing per answer, and score between 12 and 30.
OpenAIAnthropicGoogleMoonshotAlibabaxAIMistralMetaZhipuDeepSeekMiniMaxStepFunXiaomiByteDanceTencentPerceptronOpenBMBLiquidopen dot: runs locally
Higher is better, further left is cheaper. Lines show the 95% interval. Dot colour shows the maker. Open dots run on our own graphics card and cost nothing per answer, so they sit in the $0 column. The dashed line joins the models with the best score for their price: nothing cheaper scores higher.
How to read the scores
A score of 0 means no better than guessing. A score of 100 means always right. We correct for guessing because tests differ: on a yes-or-no question a model is right half the time by luck, while on “what time does the clock show?” luck gets almost nothing. Without the correction, 75% on a two-choice test and 75% on a free-answer test would look equal, though the second is far harder. The formula is (accuracy − chance) ÷ (1 − chance), and it never goes below 0. Chance comes from the number of answer choices on each question; free answers such as a time, a number or a code have chance 0. Where we show raw accuracy, we say so.
Every score has a margin of error, shown on the leaderboard. Some models were run on an evenly spread sample of each test. Twelve models were tested on a smaller sample still, so their margins are the widest. Read small gaps with care. A one-point difference is a tie.
How we test
Most test pictures are built by script, in a 3D renderer or a drawing program, with known answers, then made harder in steps. Each test has difficulty levels from easy to hard. We ask every model the same question in the same format. Answers are checked automatically against the known answer. We reviewed the tests and dropped items whose key could be wrong, so no model is marked down for a faulty question. Real test items are never published; any pictures on this site are separate public examples. Some models take only a limited number of pictures per question, so two of them could not take the video tests; their scores average the tests they could take. The full method is on the method page.
What is next
We will add models as they come out and run more of each test on the sampled models, starting with the 12 on the smallest sample, so their margins shrink. Tell us which model or skill to test next.
Tested between 22 October and 10 November 2026. Each model’s own dates are on the leaderboard.