EyeBench v3: a vision exam for frontier models (EyeBench)
We show this table for reference; we do not rank on it.
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.
Accuracy on EyeBench — EyeBench data of November 18, 2026
We mirror the published accuracy view for EyeBench. GPT-6.2 Astra leads the public snapshot at 99%, followed by GPT-6.2 Sol (98%) and GPT-6 Astra (95%). We do not use these results to rank models overall.
GPT-6.2 Astra
OpenAI
max reasoning
99/100 correct (±1.0) · $19.17 for the run
GPT-6.2 Sol
OpenAI
max reasoning
98/100 correct (±1.4) · $5.02 for the run
GPT-6 Astra
OpenAI
max reasoning
95/100 correct (±2.2) · $25.71 for the run
33 modelsMultimodal & GroundedCurrentDisplay onlyUpdated EyeBench data of November 18, 2026
Accuracy table (33 models)
| Rank | Model / configuration | Score | Parameters (B) | Open / closed |
|---|---|---|---|---|
| 1 | GPT-6.2 AstraOpenAImax reasoning99/100 correct (±1.0) · $19.17 for the run | 99% | Not reported | Closed |
| 2 | GPT-6.2 SolOpenAImax reasoning98/100 correct (±1.4) · $5.02 for the run | 98% | Not reported | Closed |
| 3 | GPT-6 AstraOpenAImax reasoning95/100 correct (±2.2) · $25.71 for the run | 95% | Not reported | Closed |
| 4 | GPT-6.1 TerraOpenAImax reasoning94/100 correct (±2.4) · $7.03 for the run | 94% | Not reported | Closed |
| 5 | GPT-6.1 SolOpenAImax reasoning93/100 correct (±2.5) · $6.25 for the run | 93% | Not reported | Closed |
| 6 | Claude Fable 5.2Anthropicmax reasoning86/100 correct (±3.5) · 1 unanswered · $65.56 for the run | 86% | Not reported | Closed |
| 7 | Claude Opus 5.5Anthropicmax reasoning78/100 correct (±4.1) · $47.94 for the run | 78% | Not reported | Closed |
| 8 | Claude Sonnet 5.5Anthropicmax reasoning71/100 correct (±4.5) · 2 unanswered · $41.43 for the run | 71% | Not reported | Closed |
| 9 | Gemini 4 ArgonGooglehigh reasoning66/100 correct (±4.7) · $23.75 for the run | 66% | Not reported | Closed |
| 10 | GPT-6 SolOpenAImax reasoning59/100 correct (±4.9) · $7.89 for the run | 59% | Not reported | Closed |
| 11 | GPT-5.6 SolOpenAImax reasoning58/100 correct (±4.9) · 8 unanswered · $54.79 for the run | 58% | Not reported | Closed |
| 12 | Claude Haiku 5.5Anthropicmax reasoning50/100 correct (±5.0) · 19 unanswered · $2.73 for the run | 50% | Not reported | Closed |
| 13 | GPT-5.6 TerraOpenAImax reasoning42/100 correct (±4.9) · 1 unanswered · $44.80 for the run | 42% | Not reported | Closed |
| 14 | GPT-5.5OpenAIxhigh reasoning41/100 correct (±4.9) · $44.65 for the run | 41% | Not reported | Closed |
| 15 | GPT-5.5 ProOpenAImedium reasoning41/100 correct (±4.9) · $243.56 for the run | 41% | Not reported | Closed |
| 16 | Claude Fable 5.1Anthropicmax reasoning40/100 correct (±4.9) · 5 unanswered · $165.97 for the run | 40% | Not reported | Closed |
| 17 | Qwen3.8 MaxAlibabaxhigh reasoning39/100 correct (±4.9) · $4.88 for the run | 39% | Not reported | Open |
| 18 | Muse Spark 1.3Metaxhigh reasoning37/100 correct (±4.8) · 23 unanswered · $4.22 for the run | 37% | Not reported | Closed |
| 19 | Muse Spark 1.2Metaxhigh reasoning34/100 correct (±4.7) · $8.69 for the run | 34% | Not reported | Closed |
| 20 | Gemini 3.8 FlashGooglehigh reasoning31/100 correct (±4.6) · $9.87 for the run | 31% | Not reported | Closed |
| 21 | GPT-5.6 LunaOpenAImax reasoning30/100 correct (±4.6) · 3 unanswered · $4.19 for the run | 30% | Not reported | Closed |
| 22 | GPT-6 LunaOpenAImax reasoning30/100 correct (±4.6) · 1 unanswered · $0.53 for the run | 30% | Not reported | Closed |
| 23 | Kimi K3Moonshot AImax reasoning29/100 correct (±4.5) · 18 unanswered · $25.66 for the run | 29% | Not reported | Open |
| 24 | Gemini 3.1 ProGooglehigh reasoning28/100 correct (±4.5) · 3 unanswered · $21.88 for the run | 28% | Not reported | Closed |
| 25 | Claude Fable 5Anthropicmax reasoning27/100 correct (±4.4) · $24.87 for the run | 27% | Not reported | Closed |
| 26 | Gemini 3.7 FlashGooglehigh reasoning27/100 correct (±4.4) · $3.27 for the run | 27% | Not reported | Closed |
| 27 | Claude Opus 5Anthropicmax reasoning26/100 correct (±4.4) · 1 unanswered · $24.15 for the run | 26% | Not reported | Closed |
| 28 | Claude Sonnet 5Anthropicmax reasoning21/100 correct (±4.1) · 1 unanswered · $11.45 for the run | 21% | Not reported | Closed |
| 29 | GLM-5.3-FlashZ.AImax reasoning21/100 correct (±4.1) · 5 unanswered · $0.31 for the run | 21% | Not reported | Open |
| 30 | Grok 4.6xAIxhigh reasoning19/100 correct (±3.9) · $0.92 for the run | 19% | Not reported | Closed |
| 31 | Mistral Large 4Mistralnone reasoning15/100 correct (±3.6) · 7 unanswered · $3.97 for the run | 15% | Not reported | Open |
| 32 | DeepSeek V4 Flash Vision ExpDeepSeekmax reasoning13/100 correct (±3.4) · $1.42 for the run | 13% | Not reported | Unknown |
| 33 | InklingThinking Machines Labxhigh reasoning9/100 correct (±2.9) · $2.31 for the run | 9% | Not reported | Open |
How to read this leaderboard
Higher is better. GPT-6.2 Astra leads with 99 of 100, ahead of GPT-6.2 Sol (98), GPT-6 Astra (95), GPT-6.1 Terra (94) and GPT-6.1 Sol (93). The best model from another lab is Claude Fable 5.2 at 86, then Claude Opus 5.5 at 78 and Gemini 4 Argon at 66. Spot the difference separates the field most: only the two GPT-6.2 models reach 90% on it, while Fable 5.2 gets 50%. Gemini 4 Argon is perfect on graphs and patterns but solves 15% of mazes. Cost varies widely for similar scores: GPT-6.2 Sol's run cost about $5, against $19 for GPT-6.2 Astra and $66 for Fable 5.2.
Operator receipt: 33 sourced rows are currently displayable on this page; the leading published row is GPT-6.2 Astra at 99%.
Honest limit: One hundred questions in five task types is a small test: one question is a point of accuracy, and the standard error near the middle of the scale is about 5 points. The images are synthetic puzzles, so the benchmark measures fine-grained counting and tracing, not photos, documents or charts. Each model is shown at its highest measured reasoning effort, and effort names are set by each provider, so they do not line up across vendors. Failed requests count as wrong, so an API outage during a run lowers a score. EyeBench is a one-person project; its questions are not published, and BenchLM cannot re-run them. BenchLM shows EyeBench as display only; it does not enter overall or category rankings.
How EyeBench is shown here
We mirror the EyeBench v3 leaderboard by Adonis Singh. Each model gets 100 generated images, 20 in each of 5 task types: 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. Every answer is a single number, marked right or wrong.
A question with no parsed answer, an invalid output or a failed request counts as wrong. Each model is listed once, at the highest reasoning effort it was run at, as on the source board. Effort names are set by each provider and do not line up across vendors. Cost is the run's tokens at the model's API prices. EyeBench is display only on BenchLM.
Snapshot
The published EyeBench snapshot places GPT-6.2 Astra first at 99%. The third row is 4 points behind. The broader top-10 range is 40 points, so the table still separates the published systems.
33 models have been evaluated on EyeBench. The benchmark falls in the Multimodal & Grounded category. We keep external benchmark mirrors separate from the weighted global scoring system, so these results remain source-specific evidence. EyeBench is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
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
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.
Freshness and provenance
Version
EyeBench 2026
Refresh cadence
Quarterly
Staleness state
Current
Question availability
Public benchmark set
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
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.
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 7,000+ subscribers.
One email each week. Unsubscribe anytime.