# VISTA > Source: Benchize (https://benchize.com) — https://benchize.com/benchmarks/scale%3Avisual_language_understanding. When citing or reusing this content, credit "Benchize" and link to https://benchize.com/benchmarks/scale%3Avisual_language_understanding. The growing capabilities of multimodal language models have created two critical challenges in their evaluation. First, as these models become more prevalent, their potential exposure to visual test data during training compromises benchmark integrity and evaluation reliability. Second, traditional evaluation approaches that focus solely on answer accuracy provide limited insight into models' visual reasoning processes, particularly when dealing with open-ended responses where simple answer normalization might prove inadequate. To address these challenges, we present VISTA, a novel multimodal benchmark designed to evaluate complex visual-language understanding. VISTA, short for Visual Task Assessment, tests models' abilities across both natural images and graphical content, requiring the integration of multiple perception skills - from OCR and spatial understanding to object recognition - while engaging broader reasoning capabilities in logic, calculation, and common sense. Each task is evaluated through a structured rubric of yes/no questions that decompose responses into specific testable conditions. This design, combined with the requirement that each question has challenged at least one prominent language model, ensures VISTA serves as a demanding testbed for visual reasoning capabilities. Canonical URL: https://benchize.com/benchmarks/scale%3Avisual_language_understanding Last updated: Wed, 21 Oct 2026 18:05:10 GMT Reproduction and quotation must name Benchize, link to the canonical HTML URL, and retain individual source citations. The growing capabilities of multimodal language models have created two critical challenges in their evaluation. First, as these models become more prevalent, their potential exposure to visual test data during training compromises benchmark integrity and evaluation reliability. Second, traditional evaluation approaches that focus solely on answer accuracy provide limited insight into models' visual reasoning processes, particularly when dealing with open-ended responses where simple answer normalization might prove inadequate. To address these challenges, we present VISTA, a novel multimodal benchmark designed to evaluate complex visual-language understanding. VISTA, short for Visual Task Assessment, tests models' abilities across both natural images and graphical content, requiring the integration of multiple perception skills - from OCR and spatial understanding to object recognition - while engaging broader reasoning capabilities in logic, calculation, and common sense. Each task is evaluated through a structured rubric of yes/no questions that decompose responses into specific testable conditions. This design, combined with the requirement that each question has challenged at least one prominent language model, ensures VISTA serves as a demanding testbed for visual reasoning capabilities. Multimodal · Models tested: 16 · Score publishers: [scale](https://labs.scale.com/leaderboard/visual_language_understanding) ## Model results | # | Model | Score | ± | | | --- | --- | --- | --- | --- | | 1 | [Gemini 2.5 Pro Preview 06-05](https://benchize.com/models/gemini-2-5-pro-preview.md) | 54.63source score | CI 54.08–55.18 | | | ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `score` \| 54.63 \| source_score \| reported \| Higher is better \| ### Configuration Reasoning effort unknown Model configuration gemini-2.5-pro-preview-06-05 sourceModelVersion Not reported thinkingEnabled Not reported ### Reported statistics Uncertainty Confidence interval: 54.08–55.18 contaminationWarning Not reported rank 2 reportedConfidenceHalfWidth 0.55 ### Source [Original source](https://labs.scale.com/leaderboard/visual_language_understanding) Source scale Original model label gemini-2.5-pro-preview-06-05 Observed 2026-10-21T16:35:38.974Z Evidence SHA-256 `469e2d64808796dc35f1a8b8b0abe7daefeb7c5ee06fa9b6ba8b34c42d4b19b5` Source locator `/rows/1` | | | | | | 2 | [Claude Opus 4](https://benchize.com/models/claude-opus-4-thinking.md)Thinking | 46.96source score | CI 46.01–47.91 | | | ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `score` \| 46.96 \| source_score \| reported \| Higher is better \| ### Configuration Reasoning effort unknown Model configuration Claude Opus 4 (Thinking) sourceModelVersion Not reported thinkingEnabled true ### Reported statistics Uncertainty Confidence interval: 46.01–47.91 contaminationWarning Not reported rank 21 reportedConfidenceHalfWidth 0.95 ### Source [Original source](https://labs.scale.com/leaderboard/visual_language_understanding) Source scale Original model label Claude Opus 4 (Thinking) Observed 2026-10-21T16:35:38.974Z Evidence SHA-256 `469e2d64808796dc35f1a8b8b0abe7daefeb7c5ee06fa9b6ba8b34c42d4b19b5` Source locator `/rows/20` | | | | | | 3 | [Gemini 3.1 Flash Lite Preview](https://benchize.com/models/gemini-3-1-flash-lite-preview.md) | 46.93source score | CI 44.19–49.67 | | | ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `score` \| 46.93 \| source_score \| reported \| Higher is better \| ### Configuration Reasoning effort unknown Model configuration gemini-3.1-flash-lite-preview sourceModelVersion Not reported thinkingEnabled Not reported ### Reported statistics Uncertainty Confidence interval: 44.19–49.67 contaminationWarning Not reported rank 22 reportedConfidenceHalfWidth 2.74 ### Source [Original source](https://labs.scale.com/leaderboard/visual_language_understanding) Source scale Original model label gemini-3.1-flash-lite-preview Observed 2026-10-21T16:35:38.974Z Evidence SHA-256 `469e2d64808796dc35f1a8b8b0abe7daefeb7c5ee06fa9b6ba8b34c42d4b19b5` Source locator `/rows/21` | | | | | | 4 | [Claude Sonnet 4](https://benchize.com/models/claude-sonnet-4-thinking.md)Thinking | 45.49source score | CI 45.28–45.7 | | | ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `score` \| 45.49 \| source_score \| reported \| Higher is better \| ### Configuration Reasoning effort unknown Model configuration Claude Sonnet 4 (Thinking) sourceModelVersion Not reported thinkingEnabled true ### Reported statistics Uncertainty Confidence interval: 45.28–45.7 contaminationWarning Not reported rank 27 reportedConfidenceHalfWidth 0.21 ### Source [Original source](https://labs.scale.com/leaderboard/visual_language_understanding) Source scale Original model label Claude Sonnet 4 (Thinking) Observed 2026-10-21T16:35:38.974Z Evidence SHA-256 `469e2d64808796dc35f1a8b8b0abe7daefeb7c5ee06fa9b6ba8b34c42d4b19b5` Source locator `/rows/26` | | | | | | 5 | [Claude Opus 4.6](https://benchize.com/models/claude-opus-4-6-off.md)Non-thinking | 45.48source score | CI 45.04–45.92 | | | ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `score` \| 45.48 \| source_score \| reported \| Higher is better \| ### Configuration Reasoning effort non-thinking Model configuration claude-opus-4-6 (Non-Thinking) sourceModelVersion Not reported thinkingEnabled false ### Reported statistics Uncertainty Confidence interval: 45.04–45.92 contaminationWarning Not reported rank 28 reportedConfidenceHalfWidth 0.44 ### Source [Original source](https://labs.scale.com/leaderboard/visual_language_understanding) Source scale Original model label claude-opus-4-6 (Non-Thinking) Observed 2026-10-21T16:35:38.974Z Evidence SHA-256 `469e2d64808796dc35f1a8b8b0abe7daefeb7c5ee06fa9b6ba8b34c42d4b19b5` Source locator `/rows/27` | | | | | | 6 | [GPT-4.1](https://benchize.com/models/gpt-4-1.md) | 45.34source score | CI 44.43–46.25 | | | ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `score` \| 45.34 \| source_score \| reported \| Higher is better \| ### Configuration Reasoning effort unknown Model configuration GPT-4.1 sourceModelVersion Not reported thinkingEnabled Not reported ### Reported statistics Uncertainty Confidence interval: 44.43–46.25 contaminationWarning Not reported rank 29 reportedConfidenceHalfWidth 0.91 ### Source [Original source](https://labs.scale.com/leaderboard/visual_language_understanding) Source scale Original model label GPT-4.1 Observed 2026-10-21T16:35:38.974Z Evidence SHA-256 `469e2d64808796dc35f1a8b8b0abe7daefeb7c5ee06fa9b6ba8b34c42d4b19b5` Source locator `/rows/28` | | | | | | 7 | [Claude Opus 4](https://benchize.com/models/claude-opus-4.md) | 43.53source score | CI 42.29–44.77 | | | ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `score` \| 43.53 \| source_score \| reported \| Higher is better \| ### Configuration Reasoning effort unknown Model configuration Claude Opus 4 sourceModelVersion Not reported thinkingEnabled Not reported ### Reported statistics Uncertainty Confidence interval: 42.29–44.77 contaminationWarning Not reported rank 35 reportedConfidenceHalfWidth 1.24 ### Source [Original source](https://labs.scale.com/leaderboard/visual_language_understanding) Source scale Original model label Claude Opus 4 Observed 2026-10-21T16:35:38.974Z Evidence SHA-256 `469e2d64808796dc35f1a8b8b0abe7daefeb7c5ee06fa9b6ba8b34c42d4b19b5` Source locator `/rows/34` | | | | | | 8 | [Claude Sonnet 4](https://benchize.com/models/claude-sonnet-4.md) | 43.21source score | CI 42.69–43.73 | | | ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `score` \| 43.21 \| source_score \| reported \| Higher is better \| ### Configuration Reasoning effort unknown Model configuration Claude Sonnet 4 sourceModelVersion Not reported thinkingEnabled Not reported ### Reported statistics Uncertainty Confidence interval: 42.69–43.73 contaminationWarning Not reported rank 37 reportedConfidenceHalfWidth 0.52 ### Source [Original source](https://labs.scale.com/leaderboard/visual_language_understanding) Source scale Original model label Claude Sonnet 4 Observed 2026-10-21T16:35:38.974Z Evidence SHA-256 `469e2d64808796dc35f1a8b8b0abe7daefeb7c5ee06fa9b6ba8b34c42d4b19b5` Source locator `/rows/36` | | | | | | 9 | [GPT-4.1 Mini](https://benchize.com/models/gpt-4-1-mini.md) | 41.14source score | CI 40.56–41.72 | | | ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `score` \| 41.14 \| source_score \| reported \| Higher is better \| ### Configuration Reasoning effort unknown Model configuration GPT-4.1 mini sourceModelVersion Not reported thinkingEnabled Not reported ### Reported statistics Uncertainty Confidence interval: 40.56–41.72 contaminationWarning Not reported rank 41 reportedConfidenceHalfWidth 0.58 ### Source [Original source](https://labs.scale.com/leaderboard/visual_language_understanding) Source scale Original model label GPT-4.1 mini Observed 2026-10-21T16:35:38.974Z Evidence SHA-256 `469e2d64808796dc35f1a8b8b0abe7daefeb7c5ee06fa9b6ba8b34c42d4b19b5` Source locator `/rows/40` | | | | | | 10 | [Llama 4 Maverick](https://benchize.com/models/llama-4-maverick.md) | 38.33source score | CI 37.78–38.88 | | | ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `score` \| 38.33 \| source_score \| reported \| Higher is better \| ### Configuration Reasoning effort unknown Model configuration Llama 4 Maverick sourceModelVersion Not reported thinkingEnabled Not reported ### Reported statistics Uncertainty Confidence interval: 37.78–38.88 contaminationWarning Not reported rank 46 reportedConfidenceHalfWidth 0.55 ### Source [Original source](https://labs.scale.com/leaderboard/visual_language_understanding) Source scale Original model label Llama 4 Maverick Observed 2026-10-21T16:35:38.974Z Evidence SHA-256 `469e2d64808796dc35f1a8b8b0abe7daefeb7c5ee06fa9b6ba8b34c42d4b19b5` Source locator `/rows/45` | | | | | | 11 | [Gemini 1.5 Pro](https://benchize.com/models/gemini-1-5-pro.md) | 37.07source score | CI 35.73–38.41 | | | ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `score` \| 37.07 \| source_score \| reported \| Higher is better \| ### Configuration Reasoning effort unknown Model configuration Gemini 1.5 Pro sourceModelVersion Not reported thinkingEnabled Not reported ### Reported statistics Uncertainty Confidence interval: 35.73–38.41 contaminationWarning Not reported rank 48 reportedConfidenceHalfWidth 1.34 ### Source [Original source](https://labs.scale.com/leaderboard/visual_language_understanding) Source scale Original model label Gemini 1.5 Pro Observed 2026-10-21T16:35:38.974Z Evidence SHA-256 `469e2d64808796dc35f1a8b8b0abe7daefeb7c5ee06fa9b6ba8b34c42d4b19b5` Source locator `/rows/47` | | | | | | 12 | [Mistral Medium 3](https://benchize.com/models/mistral-medium-3.md) | 34.59source score | CI 33.47–35.71 | | | ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `score` \| 34.59 \| source_score \| reported \| Higher is better \| ### Configuration Reasoning effort unknown Model configuration Mistral Medium 3 sourceModelVersion Not reported thinkingEnabled Not reported ### Reported statistics Uncertainty Confidence interval: 33.47–35.71 contaminationWarning Not reported rank 51 reportedConfidenceHalfWidth 1.12 ### Source [Original source](https://labs.scale.com/leaderboard/visual_language_understanding) Source scale Original model label Mistral Medium 3 Observed 2026-10-21T16:35:38.974Z Evidence SHA-256 `469e2d64808796dc35f1a8b8b0abe7daefeb7c5ee06fa9b6ba8b34c42d4b19b5` Source locator `/rows/50` | | | | | | 13 | [Gemini 1.5 Flash 002](https://benchize.com/models/gemini-1-5-flash-002.md) | 34.03source score | CI 32.62–35.44 | | | ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `score` \| 34.03 \| source_score \| reported \| Higher is better \| ### Configuration Reasoning effort unknown Model configuration Gemini 1.5 Flash 002 sourceModelVersion Not reported thinkingEnabled Not reported ### Reported statistics Uncertainty Confidence interval: 32.62–35.44 contaminationWarning Not reported rank 52 reportedConfidenceHalfWidth 1.41 ### Source [Original source](https://labs.scale.com/leaderboard/visual_language_understanding) Source scale Original model label Gemini 1.5 Flash 002 Observed 2026-10-21T16:35:38.974Z Evidence SHA-256 `469e2d64808796dc35f1a8b8b0abe7daefeb7c5ee06fa9b6ba8b34c42d4b19b5` Source locator `/rows/51` | | | | | | 14 | [Claude 3 Opus](https://benchize.com/models/claude-3-opus.md) | 27.82source score | CI 27.27–28.37 | | | ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `score` \| 27.82 \| source_score \| reported \| Higher is better \| ### Configuration Reasoning effort unknown Model configuration Claude 3 Opus sourceModelVersion Not reported thinkingEnabled Not reported ### Reported statistics Uncertainty Confidence interval: 27.27–28.37 contaminationWarning Not reported rank 56 reportedConfidenceHalfWidth 0.55 ### Source [Original source](https://labs.scale.com/leaderboard/visual_language_understanding) Source scale Original model label Claude 3 Opus Observed 2026-10-21T16:35:38.974Z Evidence SHA-256 `469e2d64808796dc35f1a8b8b0abe7daefeb7c5ee06fa9b6ba8b34c42d4b19b5` Source locator `/rows/55` | | | | | | 15 | [GPT-4.1 Nano](https://benchize.com/models/gpt-4-1-nano.md) | 26.55source score | CI 26.2–26.9 | | | ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `score` \| 26.55 \| source_score \| reported \| Higher is better \| ### Configuration Reasoning effort unknown Model configuration GPT-4.1 nano sourceModelVersion Not reported thinkingEnabled Not reported ### Reported statistics Uncertainty Confidence interval: 26.2–26.9 contaminationWarning Not reported rank 57 reportedConfidenceHalfWidth 0.35 ### Source [Original source](https://labs.scale.com/leaderboard/visual_language_understanding) Source scale Original model label GPT-4.1 nano Observed 2026-10-21T16:35:38.974Z Evidence SHA-256 `469e2d64808796dc35f1a8b8b0abe7daefeb7c5ee06fa9b6ba8b34c42d4b19b5` Source locator `/rows/56` | | | | | | 16 | [Llama 3.2 11B Vision Instruct](https://benchize.com/models/llama-3-2-11b-vision-instruct.md) | 20.47source score | CI 20.32–20.62 | | | ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `score` \| 20.47 \| source_score \| reported \| Higher is better \| ### Configuration Reasoning effort unknown Model configuration Llama 3.2 11B Vision-Instruct sourceModelVersion Not reported thinkingEnabled Not reported ### Reported statistics Uncertainty Confidence interval: 20.32–20.62 contaminationWarning Not reported rank 62 reportedConfidenceHalfWidth 0.15 ### Source [Original source](https://labs.scale.com/leaderboard/visual_language_understanding) Source scale Original model label Llama 3.2 11B Vision-Instruct Observed 2026-10-21T16:35:38.974Z Evidence SHA-256 `469e2d64808796dc35f1a8b8b0abe7daefeb7c5ee06fa9b6ba8b34c42d4b19b5` Source locator `/rows/61` | | | | | --- Cite as: VISTA — Benchize, https://benchize.com/benchmarks/scale%3Avisual_language_understanding (retrieved 2026-10-21)