# AI Vision Index

> Independent, private tests of how well AI models see. The tests cover reading clocks, meters, charts and labels, counting, space and motion, spotting changes and errors, following cables and pipes, and not being fooled. We make the test images ourselves and never publish them. Every question has one exact right answer, which a script checks, and every result is dated.

Results updated 10 Nov 2026. 36 models on 29 tests in 6 categories.

## Overall ranking

Score out of 100, corrected for guessing (0 = no better than guessing). Test score = (accuracy - guess level) / (1 - guess level); category = mean of its tests; overall = mean of the six category scores. With 95% interval. Cost = US$ per 1,000 answers at each provider's standard price ($0 for models that run locally).

- Rank 1: GPT-6 Astra (OpenAI, API): 83.7 (95% interval 81.7–85.7), $38.32 per 1,000 answers, tested 6 Nov 2026
- Rank 2–3 (tied): GPT-6.1 Sol (OpenAI, API): 81.3 (95% interval 80.0–82.5), $7.94 per 1,000 answers, tested 6 Nov 2026
- Rank 2–3 (tied): Claude Fable 5.2 (Anthropic, API): 80.7 (95% interval 78.6–82.8), $53.10 per 1,000 answers, tested 10 Nov 2026
- Rank 4–5 (tied): Gemini 4 Argon (Google, API): 77.2 (95% interval 75.9–78.5), $6.52 per 1,000 answers, tested 6 Nov 2026
- Rank 4–5 (tied): Claude Opus 5.5 (Anthropic, API): 76.5 (95% interval 74.8–78.3), $26.16 per 1,000 answers, tested 6 Nov 2026
- Rank 6–9 (tied): Claude Sonnet 5.5 (Anthropic, API): 71.0 (95% interval 68.8–73.1), $11.60 per 1,000 answers, tested 6 Nov 2026
- Rank 6–9 (tied): Kimi K3.1 Pro (Moonshot, API): 70.6 (95% interval 66.6–75.0), $71.79 per 1,000 answers, tested 8 Nov 2026
- Rank 6–9 (tied): Qwen3.8 Max (Alibaba, API): 70.6 (95% interval 66.1–75.2), $27.40 per 1,000 answers, tested 8 Nov 2026
- Rank 6–9 (tied): Claude Fable 5.1 (Anthropic, API): 69.6 (95% interval 67.5–71.7), $87.95 per 1,000 answers, tested 6 Nov 2026
- Rank 10–18 (tied): Claude Haiku 5.5 (Anthropic, API): 63.5 (95% interval 62.1–65.0), $1.01 per 1,000 answers, tested 6 Nov 2026
- Rank 10–18 (tied): Gemini 3.8 Flash (Google, API): 61.9 (95% interval 60.2–63.7), $12.80 per 1,000 answers, tested 7 Nov 2026
- Rank 10–18 (tied): GPT-6 Luna (OpenAI, API): 60.7 (95% interval 59.3–62.2), $0.74 per 1,000 answers, tested 6 Nov 2026
- Rank 10–18 (tied): Qwen3.8-27B (Alibaba, API): 59.0 (95% interval 51.8–68.0), $7.75 per 1,000 answers, tested 7 Nov 2026
- Rank 10–18 (tied): Qwen3.8 Flash (Alibaba, API): 56.4 (95% interval 55.0–57.8), $3.04 per 1,000 answers, tested 7 Nov 2026
- Rank 10–18 (tied): Grok 4.7 (xAI, API): 54.9 (95% interval 48.1–64.1), $27.40 per 1,000 answers, tested 7 Nov 2026
- Rank 10–18 (tied): Mistral Large 4 (Mistral, API): 52.7 (95% interval 46.7–61.7), $9.59 per 1,000 answers, tested 7 Nov 2026
- Rank 10–18 (tied): Muse Spark (open) (Meta, API): 47.6 (95% interval 40.3–57.3), $19.12 per 1,000 answers, tested 7 Nov 2026
- Rank 10–18 (tied): GLM-5V-Turbo (Zhipu, API): 46.8 (95% interval 44.6–49.5), $12.50 per 1,000 answers, tested 6 Nov 2026
- Rank 19–29 (tied): DeepSeek V4.1 Flash (DeepSeek, API): 44.4 (95% interval 43.2–45.8), $1.43 per 1,000 answers, tested 6 Nov 2026
- Rank 19–29 (tied): GLM-5.3 Flash (Zhipu, API): 43.2 (95% interval 41.7–44.7), $1.40 per 1,000 answers, tested 6 Nov 2026
- Rank 19–29 (tied): Qwen3.6-35B-A3B (Alibaba, API): 43.0 (95% interval 36.9–52.4), $3.77 per 1,000 answers, tested 7 Nov 2026
- Rank 19–29 (tied): Qwen3-VL-235B (Alibaba, API): 41.2 (95% interval 35.0–50.2), $7.96 per 1,000 answers, tested 7 Nov 2026
- Rank 19–29 (tied): MiniMax M3 (MiniMax, API): 40.9 (95% interval 38.7–43.6), $3.53 per 1,000 answers, tested 6 Nov 2026
- Rank 19–29 (tied): Step-5 (preview) (StepFun, API): 40.7 (95% interval 33.2–50.3), $12.50 per 1,000 answers, tested 7 Nov 2026
- Rank 19–29 (tied): MiMo v2.6 Pro (Xiaomi, API): 40.3 (95% interval 38.6–42.6), $4.60 per 1,000 answers, tested 7 Nov 2026
- Rank 19–29 (tied): Gemini 3.5 Flash-Lite (Google, API): 40.3 (95% interval 38.7–41.8), $0.34 per 1,000 answers, tested 6 Nov 2026
- Rank 19–29 (tied): Gemma 4 31B (Google, API): 38.8 (95% interval 33.4–46.6), $1.94 per 1,000 answers, tested 7 Nov 2026
- Rank 19–29 (tied): Seed 2.1 Turbo (ByteDance, API): 38.3 (95% interval 32.3–46.5), $10.85 per 1,000 answers, tested 7 Nov 2026
- Rank 19–29 (tied): Hunyuan HY4 (Tencent, API): 37.9 (95% interval 31.2–46.5), $5.71 per 1,000 answers, tested 7 Nov 2026
- Rank 30–31 (tied): Qwen3.5-9B (Alibaba, runs locally): 29.5 (95% interval 28.0–31.1), $0 per 1,000 answers (runs locally), tested 6 Nov 2026
- Rank 30–31 (tied): Perceptron Mk1.5 (Perceptron, API): 28.8 (95% interval 27.4–30.4), $0.20 per 1,000 answers, tested 6 Nov 2026; cannot take the 2 video tests
- Rank 32: Qwen3-VL-8B (Alibaba, runs locally): 26.0 (95% interval 24.8–27.6), $0 per 1,000 answers (runs locally), tested 6 Nov 2026
- Rank 33: MiniCPM-V 4.5 (OpenBMB, runs locally): 17.1 (95% interval 15.9–18.7), $0 per 1,000 answers (runs locally), tested 6 Nov 2026
- Rank 34–36 (tied): Gemma 4 E4B (Google, runs locally): 13.7 (95% interval 12.5–15.2), $0 per 1,000 answers (runs locally), tested 6 Nov 2026
- Rank 34–36 (tied): Mistral Small 2603 (Mistral, API): 12.5 (95% interval 11.5–14.2), $0.26 per 1,000 answers, tested 6 Nov 2026; cannot take the 2 video tests
- Rank 34–36 (tied): LFM2.5-VL 1.6B (Liquid, runs locally): 12.4 (95% interval 11.4–14.2), $0 per 1,000 answers (runs locally), tested 6 Nov 2026

## Scores by category (out of 100)

- GPT-6 Astra: Reading 88, Counting 78, Space, 3D and motion 84, Spotting changes and errors 84, Diagrams 98, Not being fooled 71
- GPT-6.1 Sol: Reading 82, Counting 76, Space, 3D and motion 80, Spotting changes and errors 81, Diagrams 96, Not being fooled 72
- Claude Fable 5.2: Reading 81, Counting 64, Space, 3D and motion 86, Spotting changes and errors 86, Diagrams 95, Not being fooled 72
- Gemini 4 Argon: Reading 80, Counting 58, Space, 3D and motion 72, Spotting changes and errors 89, Diagrams 90, Not being fooled 74
- Claude Opus 5.5: Reading 75, Counting 63, Space, 3D and motion 71, Spotting changes and errors 84, Diagrams 90, Not being fooled 76
- Claude Sonnet 5.5: Reading 69, Counting 53, Space, 3D and motion 68, Spotting changes and errors 74, Diagrams 84, Not being fooled 78
- Kimi K3.1 Pro: Reading 72, Counting 53, Space, 3D and motion 62, Spotting changes and errors 77, Diagrams 91, Not being fooled 68
- Qwen3.8 Max: Reading 67, Counting 56, Space, 3D and motion 75, Spotting changes and errors 72, Diagrams 84, Not being fooled 69
- Claude Fable 5.1: Reading 63, Counting 63, Space, 3D and motion 69, Spotting changes and errors 75, Diagrams 76, Not being fooled 71
- Claude Haiku 5.5: Reading 62, Counting 48, Space, 3D and motion 48, Spotting changes and errors 72, Diagrams 79, Not being fooled 72
- Gemini 3.8 Flash: Reading 58, Counting 42, Space, 3D and motion 66, Spotting changes and errors 85, Diagrams 68, Not being fooled 53
- GPT-6 Luna: Reading 59, Counting 35, Space, 3D and motion 54, Spotting changes and errors 73, Diagrams 71, Not being fooled 73
- Qwen3.8-27B: Reading 52, Counting 47, Space, 3D and motion 52, Spotting changes and errors 78, Diagrams 60, Not being fooled 65
- Qwen3.8 Flash: Reading 58, Counting 41, Space, 3D and motion 55, Spotting changes and errors 64, Diagrams 66, Not being fooled 55
- Grok 4.7: Reading 54, Counting 53, Space, 3D and motion 50, Spotting changes and errors 68, Diagrams 40, Not being fooled 64
- Mistral Large 4: Reading 48, Counting 40, Space, 3D and motion 38, Spotting changes and errors 67, Diagrams 67, Not being fooled 57
- Muse Spark (open): Reading 54, Counting 33, Space, 3D and motion 41, Spotting changes and errors 53, Diagrams 40, Not being fooled 65
- GLM-5V-Turbo: Reading 50, Counting 19, Space, 3D and motion 49, Spotting changes and errors 63, Diagrams 47, Not being fooled 53
- DeepSeek V4.1 Flash: Reading 41, Counting 17, Space, 3D and motion 42, Spotting changes and errors 61, Diagrams 54, Not being fooled 52
- GLM-5.3 Flash: Reading 42, Counting 16, Space, 3D and motion 39, Spotting changes and errors 56, Diagrams 49, Not being fooled 57
- Qwen3.6-35B-A3B: Reading 62, Counting 13, Space, 3D and motion 32, Spotting changes and errors 46, Diagrams 47, Not being fooled 57
- Qwen3-VL-235B: Reading 46, Counting 20, Space, 3D and motion 34, Spotting changes and errors 42, Diagrams 40, Not being fooled 65
- MiniMax M3: Reading 41, Counting 15, Space, 3D and motion 37, Spotting changes and errors 54, Diagrams 52, Not being fooled 47
- Step-5 (preview): Reading 42, Counting 20, Space, 3D and motion 32, Spotting changes and errors 46, Diagrams 47, Not being fooled 57
- MiMo v2.6 Pro: Reading 36, Counting 19, Space, 3D and motion 37, Spotting changes and errors 58, Diagrams 46, Not being fooled 46
- Gemini 3.5 Flash-Lite: Reading 40, Counting 22, Space, 3D and motion 39, Spotting changes and errors 44, Diagrams 44, Not being fooled 53
- Gemma 4 31B: Reading 36, Counting 0, Space, 3D and motion 31, Spotting changes and errors 48, Diagrams 40, Not being fooled 78
- Seed 2.1 Turbo: Reading 27, Counting 27, Space, 3D and motion 16, Spotting changes and errors 46, Diagrams 47, Not being fooled 68
- Hunyuan HY4: Reading 33, Counting 7, Space, 3D and motion 24, Spotting changes and errors 53, Diagrams 47, Not being fooled 64
- Qwen3.5-9B: Reading 36, Counting 19, Space, 3D and motion 27, Spotting changes and errors 24, Diagrams 25, Not being fooled 46
- Perceptron Mk1.5: Reading 39, Counting 11, Space, 3D and motion 36, Spotting changes and errors 16, Diagrams 24, Not being fooled 47
- Qwen3-VL-8B: Reading 33, Counting 18, Space, 3D and motion 22, Spotting changes and errors 18, Diagrams 19, Not being fooled 46
- MiniCPM-V 4.5: Reading 21, Counting 12, Space, 3D and motion 6, Spotting changes and errors 16, Diagrams 7, Not being fooled 40
- Gemma 4 E4B: Reading 16, Counting 3, Space, 3D and motion 10, Spotting changes and errors 7, Diagrams 7, Not being fooled 39
- Mistral Small 2603: Reading 15, Counting 10, Space, 3D and motion 7, Spotting changes and errors 12, Diagrams 10, Not being fooled 22
- LFM2.5-VL 1.6B: Reading 18, Counting 6, Space, 3D and motion 3, Spotting changes and errors 15, Diagrams 4, Not being fooled 29

## Video understanding

Two tests send a short clip as a series of pictures: Follow the hidden object and Counting on a conveyor. Average score 41 on the video tests against 52 on the other tests (models that took both). Some models accept only a handful of pictures per question and cannot take them.
- GPT-6 Astra: Follow the hidden object 97, Counting on a conveyor 100
- GPT-6.1 Sol: Follow the hidden object 92, Counting on a conveyor 98
- Claude Fable 5.2: Follow the hidden object 100, Counting on a conveyor 98
- Gemini 4 Argon: Follow the hidden object 52, Counting on a conveyor 83
- Claude Opus 5.5: Follow the hidden object 88, Counting on a conveyor 92
- Claude Sonnet 5.5: Follow the hidden object 70, Counting on a conveyor 53
- Kimi K3.1 Pro: Follow the hidden object 72, Counting on a conveyor 93
- Qwen3.8 Max: Follow the hidden object 72, Counting on a conveyor 60
- Claude Fable 5.1: Follow the hidden object 92, Counting on a conveyor 87
- Claude Haiku 5.5: Follow the hidden object 35, Counting on a conveyor 78
- Gemini 3.8 Flash: Follow the hidden object 56, Counting on a conveyor 41
- GPT-6 Luna: Follow the hidden object 47, Counting on a conveyor 33
- Qwen3.8-27B: Follow the hidden object 45, Counting on a conveyor 60
- Qwen3.8 Flash: Follow the hidden object 53, Counting on a conveyor 61
- Grok 4.7: Follow the hidden object 45, Counting on a conveyor 40
- Mistral Large 4: Follow the hidden object 17, Counting on a conveyor 20
- Muse Spark (open): Follow the hidden object 17, Counting on a conveyor 20
- GLM-5V-Turbo: Follow the hidden object 42, Counting on a conveyor 10
- DeepSeek V4.1 Flash: Follow the hidden object 37, Counting on a conveyor 22
- GLM-5.3 Flash: Follow the hidden object 30, Counting on a conveyor 12
- Qwen3.6-35B-A3B: Follow the hidden object 17, Counting on a conveyor 0
- Qwen3-VL-235B: Follow the hidden object 25, Counting on a conveyor 0
- MiniMax M3: Follow the hidden object 17, Counting on a conveyor 25
- Step-5 (preview): Follow the hidden object 50, Counting on a conveyor 0
- MiMo v2.6 Pro: Follow the hidden object 28, Counting on a conveyor 17
- Gemini 3.5 Flash-Lite: Follow the hidden object 24, Counting on a conveyor 18
- Gemma 4 31B: Follow the hidden object 45, Counting on a conveyor 0
- Seed 2.1 Turbo: Follow the hidden object 45, Counting on a conveyor 0
- Hunyuan HY4: Follow the hidden object 45, Counting on a conveyor 20
- Qwen3.5-9B: Follow the hidden object 0, Counting on a conveyor 28
- Perceptron Mk1.5: Follow the hidden object not supported, Counting on a conveyor not supported
- Qwen3-VL-8B: Follow the hidden object 17, Counting on a conveyor 23
- MiniCPM-V 4.5: Follow the hidden object 5, Counting on a conveyor 13
- Gemma 4 E4B: Follow the hidden object 0, Counting on a conveyor 0
- Mistral Small 2603: Follow the hidden object not supported, Counting on a conveyor not supported
- LFM2.5-VL 1.6B: Follow the hidden object 0, Counting on a conveyor 4

## The tests

- Reading: Reading a clock; Reading a meter; Eye chart; Reading a label; Reading a chart; Test strips; Adding up coins; Traffic light from a car
- Counting: Counting objects; Counting stacked goods; Counting on a conveyor
- Space, 3D and motion: Which is nearer; Which way it faces; Same or mirror image; Will it topple; Match the render; Which bay can I park in?; Follow the hidden object
- Spotting changes and errors: Spot the change; Shelf against plan; Rendering glitch; Check the invoice
- Diagrams: Follow the cable; Follow the pipe; Missing dimension
- Not being fooled: Absent objects; Saying "I can't tell"; Optical illusions; Real or AI-made

## Reports

- [We tested 36 AI vision models on private tests](https://aivisionindex.com/reports/we-tested-36-ai-vision-models/) (10 Nov 2026): AI Vision Index: 36 models from 18 makers. Who sees best, what it costs, and where every model still gets it wrong.
- [The best AI for reading gauges, labels and fine print is GPT-6 Astra, by a clear margin](https://aivisionindex.com/reports/best-ai-for-reading-dials-and-labels/) (10 Nov 2026): GPT-6 Astra scores 88 out of 100 on our eight reading tests, six points clear of the next model. On the smallest print it gets 98; Claude Opus 5.5 gets 44.
- [Can AI count past 20? Not in a cluttered picture.](https://aivisionindex.com/reports/can-ai-count-past-20/) (10 Nov 2026): Up to 10 objects, the best models count right nine times in ten or more. With 21 to 30 in a busy scene, the best gets one in five. Neat stacks and one-at-a-time video are a different story.
- [Check an invoice: yes. Count a crowd: not yet. Which picture jobs AI can do today](https://aivisionindex.com/reports/which-vision-jobs-ai-can-do/) (10 Nov 2026): On four of our 29 tests, nine or more of the 24 models tested on our larger sample score 90 out of 100 or more, and the cheapest of them cost 20 to 74 cents per 1,000 answers. On five tests, none of them reaches 80.
- [The most expensive AI model we tested is not the best](https://aivisionindex.com/reports/priciest-ai-model-not-the-best/) (10 Nov 2026): Claude Fable 5.1 costs about $88 per 1,000 answers and ranks ninth of 36. Gemini 4 Argon costs $6.52 and scores 7.6 points higher.
- [The top AI vision model can barely tell an AI-made picture from a real photo](https://aivisionindex.com/reports/ai-cant-spot-ai-pictures/) (10 Nov 2026): Gemini 4 Argon gets 81% right. GPT-6 Astra, first overall, gets 57.5% on a test where a coin flip gets 50%.
- [Video is where the best AI models pull away from the rest](https://aivisionindex.com/reports/video-is-where-the-best-pull-away/) (10 Nov 2026): On our two video tests, five models score 89 or more and 15 of 34 score under 25. For the leaders our video tests are easier than the still pictures. For nearly everyone else they are much harder.
- [AI has mostly stopped seeing things that aren't there. Optical illusions still fool a third of the models.](https://aivisionindex.com/reports/ai-stopped-imagining-but-illusions-fool/) (10 Nov 2026): On questions about absent objects, 32 of 36 models score 85 or more out of 100. On optical illusions, 11 of 36 score 3 or less.
- [The best cheap AI vision model: 61 out of 100 for 74 cents per 1,000 answers](https://aivisionindex.com/reports/best-cheap-ai-vision-model/) (10 Nov 2026): Four models cost 20 to 74 cents per 1,000 answers. Their scores run from 12 to 61. In the cheap tier, which one you pick matters more than the price.
- [Small AI models that run on your own PC score a third of the best cloud model](https://aivisionindex.com/reports/ai-on-your-own-pc/) (10 Nov 2026): Five small open models ran on one ordinary PC. The best scored 29.5 out of 100, against 83.7 for GPT-6 Astra. Google's cheapest cloud model, at 34 cents per 1,000 answers, beat all five. Bigger open models do better, but need a bigger machine.

## What's new

- 10 Nov 2026, newly tested: Claude Fable 5.2 (Anthropic) joins the leaderboard.
- 7 Nov 2026, new leaderboard: 36 models on 29 tests in six categories, scores corrected for guessing.
- 7 Nov 2026, new section: Video understanding: how the models do on two tests where things move.

## More

- [Leaderboard](https://aivisionindex.com/)
- [Scores by test and video understanding](https://aivisionindex.com/tests/)
- [Method: how we test and score](https://aivisionindex.com/method/)
- [All results as JSON](https://aivisionindex.com/data/leaderboard.json): scores, intervals, raw accuracy and guess level per test, cost and test dates. No test items.
- [中文版 / Chinese version](https://aivisionindex.com/zh/): the same pages in Simplified Chinese (https://aivisionindex.com/zh/llms.txt)

## Model pages

- [GPT-6 Astra](https://aivisionindex.com/models/gpt-6-astra/)
- [GPT-6.1 Sol](https://aivisionindex.com/models/gpt-6-1-sol/)
- [Claude Fable 5.2](https://aivisionindex.com/models/claude-fable-5-2/)
- [Gemini 4 Argon](https://aivisionindex.com/models/gemini-4-argon/)
- [Claude Opus 5.5](https://aivisionindex.com/models/claude-opus-5-5/)
- [Claude Sonnet 5.5](https://aivisionindex.com/models/claude-sonnet-5-5/)
- [Kimi K3.1 Pro](https://aivisionindex.com/models/kimi-k3-1-pro/)
- [Qwen3.8 Max](https://aivisionindex.com/models/qwen3-8-max/)
- [Claude Fable 5.1](https://aivisionindex.com/models/claude-fable-5-1/)
- [Claude Haiku 5.5](https://aivisionindex.com/models/claude-haiku-5-5/)
- [Gemini 3.8 Flash](https://aivisionindex.com/models/gemini-3-8-flash/)
- [GPT-6 Luna](https://aivisionindex.com/models/gpt-6-luna/)
- [Qwen3.8-27B](https://aivisionindex.com/models/qwen3-8-27b/)
- [Qwen3.8 Flash](https://aivisionindex.com/models/qwen3-8-flash/)
- [Grok 4.7](https://aivisionindex.com/models/grok-4-7/)
- [Mistral Large 4](https://aivisionindex.com/models/mistral-large-4/)
- [Muse Spark (open)](https://aivisionindex.com/models/muse-spark-open/)
- [GLM-5V-Turbo](https://aivisionindex.com/models/glm-5v-turbo/)
- [DeepSeek V4.1 Flash](https://aivisionindex.com/models/deepseek-v4-1-flash/)
- [GLM-5.3 Flash](https://aivisionindex.com/models/glm-5-3-flash/)
- [Qwen3.6-35B-A3B](https://aivisionindex.com/models/qwen3-6-35b-a3b/)
- [Qwen3-VL-235B](https://aivisionindex.com/models/qwen3-vl-235b/)
- [MiniMax M3](https://aivisionindex.com/models/minimax-m3/)
- [Step-5 (preview)](https://aivisionindex.com/models/step-5-preview/)
- [MiMo v2.6 Pro](https://aivisionindex.com/models/mimo-v2-6-pro/)
- [Gemini 3.5 Flash-Lite](https://aivisionindex.com/models/gemini-3-5-flash-lite/)
- [Gemma 4 31B](https://aivisionindex.com/models/gemma-4-31b/)
- [Seed 2.1 Turbo](https://aivisionindex.com/models/seed-2-1-turbo/)
- [Hunyuan HY4](https://aivisionindex.com/models/hunyuan-hy4/)
- [Qwen3.5-9B](https://aivisionindex.com/models/qwen3-5-9b/)
- [Perceptron Mk1.5](https://aivisionindex.com/models/perceptron-mk1-5/)
- [Qwen3-VL-8B](https://aivisionindex.com/models/qwen3-vl-8b/)
- [MiniCPM-V 4.5](https://aivisionindex.com/models/minicpm-v-4-5/)
- [Gemma 4 E4B](https://aivisionindex.com/models/gemma-4-e4b/)
- [Mistral Small 2603](https://aivisionindex.com/models/mistral-small-2603/)
- [LFM2.5-VL 1.6B](https://aivisionindex.com/models/lfm2-5-vl-1-6b/)

- [FAQ and glossary](https://aivisionindex.com/faq/)
- [Compare models](https://aivisionindex.com/compare/)
- Licence: the aggregate results are CC BY 4.0.