Models compared
Pick a brain. We tested them.
Every model below answered the same quiz questions inside Cloudy. Here's how fast each one was, and what 1,000 scans would cost on your own key.
| Model | Right answers | Typical time | Cost per 1,000 scans |
|---|---|---|---|
| Gemini 3.5 Flash-LiteRecommended | 14 / 14 | 3.0 s | Free with a free Gemini key$1.31 on a paid key |
| GPT-6.1 SolOpenAI | 14 / 14 | 3.2 s | $7.12 |
| GPT-6 LunaOpenAI | 14 / 14 | 4.5 s | $0.39 |
| Gemini 3.8 FlashGoogle | 14 / 14 | 5.2 s | $1.23 |
| Gemini 3.5 FlashGoogle | 14 / 14 | 5.3 s | $1.95 |
- With Gemini 3.5 Flash-Lite, Cloudy asks for a quick answer and a careful one at the same time and uses whichever is ready, so a scan usually answers in about 2 seconds.
- Finding the answer on your screen and clicking it takes about 0.2 seconds more, on your Mac.
- Models that run on your own Mac are free and fully private, but in our tests a scan took 20 to 30 seconds and was right less often. Good for privacy, not for speed.
- Claude models weren't in this round. They work in Cloudy with your Anthropic key.
How we tested: 14 multiple-choice screenshots (Kahoot, Blooket and quiz pages), one careful answer per scan, on October 4, 2026, from Taiwan. Times include sending the screenshot, so yours depend on your internet. Costs use each provider's standard price that day.
Beyond scans
Chat, code and doing things on a computer, with DeepSeek and Qwen next to the others.
| Model | Chat and code | Scans | On a computer | Per 1M tokens |
|---|---|---|---|---|
| Gemini 3.5 Flash-LiteGoogle | 8 / 82 to 4 s | 14 / 143.3 s | 5 / 510 to 38 s | Free with a free Gemini key |
| GPT-6 LunaOpenAI | 8 / 83 to 4 s | 14 / 145.9 s | 5 / 514 to 54 s | $0.10 / $0.50 |
| DeepSeek V4.1 FlashDeepSeek | 8 / 81 to 6 s | 14 / 146.3 s, slowest 60 s | 5 / 510 to 63 s | $0.30 / $1.20 |
| Qwen 3.7 PlusAlibaba Qwen | 8 / 86 to 37 s | 0 / 14empty answers | 4 / 512 to 53 s | $0.32 / $1.28 |
| Qwen 3.7 FlashAlibaba Qwen | 8 / 82 to 25 s | 0 / 14empty answers | 2 / 5one stopped by a content filter | $0.03 / $0.13 |
| Qwen 3.8 FlashAlibaba Qwen | 3 / 8answered, rate-limited | 5 / 14answered, all right | 0 / 5answered, rate-limited | $0.15 / $0.47 |
- DeepSeek V4.1 Flash is fast and cheap, but it writes a lot: a game can run past 12,000 tokens, so a big build takes about a minute.
- Both Qwen 3.7 models did well in chat and code, but gave empty answers to scans in Cloudy and struggled with multi-step work on a screen.
- Qwen 3.8 Flash was rate-limited most of the time during our test, and right whenever it did answer. Cloudy 9.6 now asks again a few times when a model is busy.
- DeepSeek and Qwen ran through OpenRouter, so speed and price depend on which host answers. Prices are list prices per million tokens, in and out. Bring your own with Settings › Models › Custom, then @ it in Team mode.
How we tested: 5 chat prompts (one in Chinese) and 3 coding tasks checked by running them, the same 14 scans as above, and 5 tasks on a test computer (look up facts, save a file from the text editor, search with a dropdown). October 8, 2026, from Taiwan, one run each, so treat it as a guide, not a promise.