Reasoning models, ranked
A reasoning model thinks in tokens before it answers, and those tokens are output tokens on your bill. The quality gain on multi-step problems is real, and so is the multiplier: we reserve roughly three times the usual output when a request goes to one of these, then charge only what the model actually used.
Best value in this category
Score out of 100: 50% cheapness, 30% capabilities, 20% context size — not a quality benchmark.
| # | Model | Context | ≈ 1,000 words | Score |
|---|---|---|---|---|
| 1 | Gemini 2.5 Flash Lite (batch) | 1,048,576 | 94 | 86.2 |
| 2 | Qwen3.7 Flash | 1,000,000 | 60 | 84.9 |
| 3 | Muse Spark 1.2 Contributor | 1,048,576 | 113 | 84.2 |
| 4 | Muse Spark 1.3 Contributor | 1,048,576 | 113 | 84.2 |
| 5 | GPT-5 Nano (batch) | 400,000 | 85 | 83.9 |
| 6 | Nex-N2-Mini | 262,144 | 47 | 82.9 |
| 7 | Ling 3.0 Flash | 262,144 | 32 | 81.1 |
| 8 | GLM Flash Latest | 1,310,720 | 116 | 78.6 |
| 9 | Gemini 2.5 Flash Lite | 1,048,576 | 189 | 78.6 |
| 10 | GLM 5.3 Flash | 1,310,720 | 123 | 78 |
| 11 | Solar Pro 4 | 524,288 | 57 | 77.2 |
| 12 | Qwen3.5-Flash | 1,000,000 | 123 | 77.1 |
| 13 | DeepSeek V4 Flash Latest | 1,310,720 | 79 | 76.8 |
| 14 | GPT-5 Nano | 400,000 | 170 | 76.4 |
| 15 | Qwen3.5-9B | 262,144 | 94 | 75.4 |
Full list of reasoning models, and what they cost with prices →