The Artificial Analysis chart
Llama 4 Maverick has the lowest latency of the models we track: Artificial Analysis measures 1.1 s to its first answer token, followed by Qwen3 235B A22B (1.3 s) and Gemini 3.1 Flash Lite (5.0 s).
Run current Claude, GPT, Gemini and open models in your agent from $19/month
See plans| Model | First answer tokenArtificial Analysis | Cost per requestour data |
|---|---|---|
| 1. Llama 4 Maverick | 1.1 s | $0.0095 |
| 2. Qwen3 235B A22BQwen3 235B A22B 2507 Instruct | 1.3 s | $0.0096 |
| 3. Gemini 3.1 Flash Lite | 5.0 s | $0.0082 |
| 4. DeepSeek V4.1 Flashmax effort | 11 s | $0.0067 |
| 5. Qwen3.7 Max | 14 s | $0.063 |
| 6. MiniMax M3 | 18 s | $0.013 |
| 7. Grok 4.20Grok 4.20 0309 v2, reasoning | 21 s | $0.068 |
| 8. DeepSeek V4 ProDeepSeek V4 Pro 0813, max effort | 22 s | $0.030 |
| 9. Gemini 3.1 ProGemini 3.1 Pro Preview | 23 s | $0.095 |
| 10. Gemini 3.8 Flashhigh effort | 24 s | $0.048 |
| 11. Step 5 Preview | 25 s | $0.025 |
| 12. MiniMax M2.1 | 31 s | $0.0065 |
| 13. GLM 5.3max effort | 32 s | $0.057 |
| 14. Kimi K2.7 Code | 33 s | $0.027 |
| 15. Muse Wondermax effort | 38 s | $0.11 |
| 16. Qwen3.7 Plus | 39 s | $0.014 |
| 17. GLM 4.7reasoning | 49 s | $0.017 |
| 18. GLM 4.6reasoning | 59 s | $0.015 |
| 19. GPT-5.3-Codexxhigh effort | 72 s | $0.055 |
| 20. DeepSeek R1 (0528)may '25 | 73 s | $0.031 |
| 21. Grok 4.7xhigh effort | 101 s | $0.13 |
| 22. GPT-6 Lunamax effort | 101 s | $0.0041 |
| 23. Kimi K2.6reasoning | 105 s | $0.024 |
| 24. GPT-5.4 Minixhigh effort | 2.1 min | $0.023 |
| 25. GPT-5.6 Terramax effort | 2.2 min | $0.084 |
| 26. GPT-5.4xhigh effort | 3.3 min | $0.10 |
| 27. Claude Fable 5.1max effort | 4.0 min | $0.24 |
| 28. Claude Haiku 5.5max effort | 5.2 min | $0.013 |
| 29. GPT-6.1 Solmax effort | 5.5 min | $0.067 |
| 30. GPT-6 Astramax effort | 6.0 min | $0.41 |
| 31. Claude Sonnet 5.5max effort | 7.3 min | $0.053 |
| 32. Claude Opus 5.5max effort | 11 min | $0.11 |
Source: Artificial Analysis (artificialanalysis.ai). Each model is measured as the variant named under it. Cost per request is ours: every one of 79,301 real coding-agent requests priced at list rates (OpenRouter's listed price for some older and open models). Methodology
| Model | First answer tokenArtificial Analysis | Cost per requestour data |
|---|---|---|
| Mistral Large 3superseded by Mistral Large 4 | 1.1 s | $0.012 |
| Gemini 3.5 Flashsuperseded by Gemini 3.8 Flash | 17 s | $0.049 |
| Claude Haiku 4.5superseded by Claude Haiku 5.5 | 22 s | $0.022 |
| Grok 4.3superseded by Grok 4.7 | 25 s | $0.068 |
| GLM 5.2superseded by GLM 5.3 | 27 s | $0.057 |
| Claude Opus 4.8superseded by Claude Opus 5.5 | 66 s | $0.16 |
| GPT-5.5superseded by GPT-6.1 Sol | 71 s | $0.21 |
| Claude Sonnet 4.6superseded by Claude Sonnet 5.5 | 101 s | $0.097 |
| GPT-5.6 Lunasuperseded by GPT-6 Luna | 117 s | $0.0084 |
| GPT-5.6 Solsuperseded by GPT-6.1 Sol | 2.1 min | $0.17 |
| Claude Fable 5superseded by Claude Fable 5.1 | 2.3 min | $0.32 |
| GPT-6 Solsuperseded by GPT-6.1 Sol | 2.5 min | $0.083 |
Source: Artificial Analysis (artificialanalysis.ai). Cost per request is ours. Methodology
Latency here is measured at the reasoning effort in each variant's name, mostly the highest. It is not a property of the model alone: lower effort, shorter prompts and a different host all change it.
Current Claude, GPT, Gemini and open models are included with Standard Compute
