# EyeBench v3: a vision exam for frontier models (EyeBench)

> Adonis Singh's visual reasoning benchmark: 100 generated images, each with one question whose answer is a single number. Models count the differences between two nearly identical panels, count closed loops in a tangle of curves, count the connected parts of a graph drawing, pick the tile that completes a 4x4 pattern, and count the turns along a maze path. Scoring is exact match.

Canonical page: https://benchlm.ai/benchmarks/eyebench

- Category: [Multimodal & Grounded](/multimodal-grounded)
- Last updated: EyeBench data of November 18, 2026

## About EyeBench

- Year: 2026
- Tasks: 100 generated images in five task types: spot the difference, closed loops, graph components, 4x4 pattern completion and maze turns (20 each)
- Format: One image and one question per request; the answer is a single integer
- Difficulty: Counting and tracing fine visual detail exactly: small differences between panels, crossing curves and long maze paths
- Paper: [EyeBench v3: a vision exam for frontier models](https://adonis-singh.github.io/eyebench/)

EyeBench v3 has 100 procedurally generated images, 20 in each of five task types: spot the difference (count the differences between two nearly identical panels), topology (count closed loops in a tangle of curves), graph (count connected components in a graph drawing), pattern (pick the option that completes a 4x4 visual matrix) and maze (count the turns along a path). Each question is one request with one image, and the answer must be a single integer. Scoring is exact match; on four pattern items the correct tile is pixel-identical to another option, and either is accepted. A question with no parsed answer, an invalid output or a failed request counts as wrong. The publisher runs each model at several reasoning efforts and lists it at the highest one measured; BenchLM shows that same run, once per model. Costs are the run's tokens at the model's API prices; caching does not apply, since every question is a single request. BenchLM reads the run-level data the EyeBench site publishes and keeps each task type's accuracy in its history.

EyeBench is currently displayed on BenchLM for reference, but it is excluded from the weighted scoring formula.

## Leaderboard (33 models)

| Rank | Model | Configuration | Creator | Score |
|------|-------|---------------|---------|-------|
| 1 | [GPT-6.2 Astra](/models/gpt-6-2-astra) | max reasoning · 99/100 correct (±1.0) · $19.17 for the run | OpenAI | 99% |
| 2 | [GPT-6.2 Sol](/models/gpt-6-2-sol) | max reasoning · 98/100 correct (±1.4) · $5.02 for the run | OpenAI | 98% |
| 3 | [GPT-6 Astra](/models/gpt-6-astra) | max reasoning · 95/100 correct (±2.2) · $25.71 for the run | OpenAI | 95% |
| 4 | [GPT-6.1 Terra](/models/gpt-6-1-terra) | max reasoning · 94/100 correct (±2.4) · $7.03 for the run | OpenAI | 94% |
| 5 | [GPT-6.1 Sol](/models/gpt-6-1-sol) | max reasoning · 93/100 correct (±2.5) · $6.25 for the run | OpenAI | 93% |
| 6 | [Claude Fable 5.2](/models/claude-fable-5-2) | max reasoning · 86/100 correct (±3.5) · 1 unanswered · $65.56 for the run | Anthropic | 86% |
| 7 | [Claude Opus 5.5](/models/claude-opus-5-5) | max reasoning · 78/100 correct (±4.1) · $47.94 for the run | Anthropic | 78% |
| 8 | [Claude Sonnet 5.5](/models/claude-sonnet-5-5) | max reasoning · 71/100 correct (±4.5) · 2 unanswered · $41.43 for the run | Anthropic | 71% |
| 9 | [Gemini 4 Argon](/models/gemini-4-argon) | high reasoning · 66/100 correct (±4.7) · $23.75 for the run | Google | 66% |
| 10 | [GPT-6 Sol](/models/gpt-6-sol) | max reasoning · 59/100 correct (±4.9) · $7.89 for the run | OpenAI | 59% |
| 11 | [GPT-5.6 Sol](/models/gpt-5-6-sol) | max reasoning · 58/100 correct (±4.9) · 8 unanswered · $54.79 for the run | OpenAI | 58% |
| 12 | [Claude Haiku 5.5](/models/claude-haiku-5-5) | max reasoning · 50/100 correct (±5.0) · 19 unanswered · $2.73 for the run | Anthropic | 50% |
| 13 | [GPT-5.6 Terra](/models/gpt-5-6-terra) | max reasoning · 42/100 correct (±4.9) · 1 unanswered · $44.80 for the run | OpenAI | 42% |
| 14 | [GPT-5.5](/models/gpt-5-5) | xhigh reasoning · 41/100 correct (±4.9) · $44.65 for the run | OpenAI | 41% |
| 15 | [GPT-5.5 Pro](/models/gpt-5-5-pro) | medium reasoning · 41/100 correct (±4.9) · $243.56 for the run | OpenAI | 41% |
| 16 | [Claude Fable 5.1](/models/claude-fable-5-1) | max reasoning · 40/100 correct (±4.9) · 5 unanswered · $165.97 for the run | Anthropic | 40% |
| 17 | [Qwen3.8 Max](/models/qwen3-8-max) | xhigh reasoning · 39/100 correct (±4.9) · $4.88 for the run | Alibaba | 39% |
| 18 | [Muse Spark 1.3](/models/muse-spark-1-3) | xhigh reasoning · 37/100 correct (±4.8) · 23 unanswered · $4.22 for the run | Meta | 37% |
| 19 | [Muse Spark 1.2](/models/muse-spark-1-2) | xhigh reasoning · 34/100 correct (±4.7) · $8.69 for the run | Meta | 34% |
| 20 | [Gemini 3.8 Flash](/models/gemini-3-8-flash) | high reasoning · 31/100 correct (±4.6) · $9.87 for the run | Google | 31% |
| 21 | [GPT-5.6 Luna](/models/gpt-5-6-luna) | max reasoning · 30/100 correct (±4.6) · 3 unanswered · $4.19 for the run | OpenAI | 30% |
| 22 | [GPT-6 Luna](/models/gpt-6-luna) | max reasoning · 30/100 correct (±4.6) · 1 unanswered · $0.53 for the run | OpenAI | 30% |
| 23 | [Kimi K3](/models/kimi-k3) | max reasoning · 29/100 correct (±4.5) · 18 unanswered · $25.66 for the run | Moonshot AI | 29% |
| 24 | [Gemini 3.1 Pro](/models/gemini-3-1-pro) | high reasoning · 28/100 correct (±4.5) · 3 unanswered · $21.88 for the run | Google | 28% |
| 25 | [Claude Fable 5](/models/claude-fable) | max reasoning · 27/100 correct (±4.4) · $24.87 for the run | Anthropic | 27% |
| 26 | [Gemini 3.7 Flash](/models/gemini-3-7-flash) | high reasoning · 27/100 correct (±4.4) · $3.27 for the run | Google | 27% |
| 27 | [Claude Opus 5](/models/claude-opus-5) | max reasoning · 26/100 correct (±4.4) · 1 unanswered · $24.15 for the run | Anthropic | 26% |
| 28 | [Claude Sonnet 5](/models/claude-sonnet-5) | max reasoning · 21/100 correct (±4.1) · 1 unanswered · $11.45 for the run | Anthropic | 21% |
| 29 | [GLM-5.3-Flash](/models/glm-5-3-flash) | max reasoning · 21/100 correct (±4.1) · 5 unanswered · $0.31 for the run | Z.AI | 21% |
| 30 | [Grok 4.6](/models/grok-4-6) | xhigh reasoning · 19/100 correct (±3.9) · $0.92 for the run | xAI | 19% |
| 31 | [Mistral Large 4](/models/mistral-large-4) | none reasoning · 15/100 correct (±3.6) · 7 unanswered · $3.97 for the run | Mistral | 15% |
| 32 | [DeepSeek V4 Flash Vision Exp](https://adonis-singh.github.io/eyebench/) | max reasoning · 13/100 correct (±3.4) · $1.42 for the run | DeepSeek | 13% |
| 33 | [Inkling](/models/inkling) | xhigh reasoning · 9/100 correct (±2.9) · $2.31 for the run | Thinking Machines Lab | 9% |

## FAQ

### What does EyeBench measure?

Whether a model can read fine visual detail exactly. Each of 100 generated images comes with one question whose answer is a number: how many differences between two panels, how many closed loops, how many connected parts of a graph, which tile completes a pattern, or how many turns a maze path takes. Only the exact number counts.

### Which model leads EyeBench?

GPT-6.2 Astra at max reasoning leads with 99 of 100, followed by GPT-6.2 Sol (98), GPT-6 Astra (95) and GPT-6.1 Terra (94). The best non-OpenAI model is Claude Fable 5.2 at 86 (EyeBench data of November 18, 2026).

### Why does BenchLM show one effort level per model on EyeBench?

EyeBench runs many models at several reasoning efforts, and its leaderboard lists each model at the highest effort measured. BenchLM shows the same run, so every model appears once.
