Speed
18 models ranked
The fastest AI models ranked by median output tokens per second (t/s). Speeds are re-measured regularly by Artificial Analysis across real API endpoints, so the leaderboard reflects typical real-world throughput.
Speeds measured by Artificial Analysis — median output tokens per second
| # | Model | ||
|---|---|---|---|
| 1 | GPT-4o mini | OpenAI | 96 t/s |
| 2 | Gemini 2.5 Flash | 93 t/s | |
| 3 | Grok 3 | xAI | 88 t/s |
| 4 | GPT-5 | OpenAI | 82 t/s |
| 5 | Claude 4 Sonnet | Anthropic | 79 t/s |
| 6 | Cursor Tab | Anysphere | 78 t/s |
| 7 | Gemini 2.5 Pro | 74 t/s | |
| 8 | Qwen3 235B | Alibaba | 71 t/s |
| 9 | GitHub Copilot | GitHub | 70 t/s |
| 10 | DeepSeek V3 | DeepSeek | 69 t/s |
| 11 | Llama 4 405B | Meta | 66 t/s |
| 12 | Codeium Windsurf | Codeium | 65 t/s |
| 13 | Mistral Large 3 | Mistral | 64 t/s |
| 14 | Claude Code | Anthropic | 63 t/s |
| 15 | Claude 4 Opus | Anthropic | 61 t/s |
| 16 | Aider | Aider | 59 t/s |
| 17 | Command R+ | Cohere | 58 t/s |
| 18 | Continue Dev | Continue | 55 t/s |
AIRank aggregates these results from the public sources listed on each page. We do not run our own evaluations, and we normalize raw results to a 0–100 scale so tables are comparable.
FAQ
Frequently asked questions
What does “tokens per second” mean?
Tokens per second (t/s) measures how many output tokens a model generates in one second — roughly, how much text it produces per second of generation. A higher number means faster responses: long answers arrive sooner and interactive tools feel snappier. The value shown is the median output speed measured across real API endpoints.
How is speed measured?
Output speeds are measured by Artificial Analysis, which re-tests model endpoints roughly every eight hours across many prompts and hardware configurations. We take the reported median output tokens per second, so the ranking reflects typical real-world throughput rather than a single best-case run.
Is a faster model always better?
Not necessarily. The fastest models are often small or heavily optimized variants that score lower on reasoning benchmarks. Speed and quality are different axes — use this leaderboard to see how quickly a model responds and the benchmark pages to see what it can actually do.