Benchmarks · speed
Output Speed (tokens/s)
Output tokens per second over a request's full duration, so time to first token and reasoning tokens count. AA averages its workloads (1k, 10k and 100k prompts, and multi-turn agent sessions) over the past 5 days, and a request that fails on the endpoint's side counts as 0. Read from the lab's own API where AA has one, otherwise the median across providers.
Artificial Analysis changed how it measures speed in Performance Benchmarking v3.0.0. Values before Nov 4, 2026 are not comparable with later ones. Their method
162 models · 298 results · data as of 2026-11-06 · artificialanalysis.ai ↗
Leaderboard
Mercury 2.5
All results
298 rows| # | Model | Effort | Output Speed (tokens/s) |
|---|---|---|---|
| 1 | Mercury 2.5 | 575.5 | |
| 2 | Gemini 3.1 Flash-Lite | high | 336.0 |
| 3 | Gemini 3.5 Flash-Lite | high | 298.8 |
| 4 | Mercury 2 | high | 286.8 |
| 5 | Gemini 2.5 Flash Lite | none | 275.0 |
| 6 | Nemotron 3.5 Lightning | 244.9 | |
| 7 | Haiku 5.5 | max | 235.6 |
| 8 | Haiku 5.5 | xhigh | 233.8 |
| 9 | o3 mini | high | 229.9 |
| 10 | Muse Spark 1.3 | xhigh | 225.9 |
| 11 | Haiku 5.5 | high | 219.1 |
| 12 | Haiku 5.5 | medium | 217.8 |
| 13 | Haiku 5.5 | low | 215.9 |
| 14 | Gemini 2.5 Flash | 207.6 | |
| 15 | Ling 3.0 Flash | 199.3 | |
| 16 | Gemini 3 Flash | 199.2 | |
| 17 | DeepSeek V4.1 Flash | none | 198.0 |
| 18 | DeepSeek V3 | none | 194.9 |
| 19 | GPT-5.4 Mini | xhigh | 190.9 |
| 20 | Muse Spark 1.1 | xhigh | 187.0 |
| 21 | GPT-5.4 Mini | medium | 185.3 |
| 22 | GLM 5.2 | none | 184.3 |
| 23 | Gemini 3.5 Flash | high | 183.9 |
| 24 | o3 mini | 183.5 | |
| 25 | Gemini 3.6 Flash | high | 182.1 |
| 26 | Gemini 2.5 Flash | none | 180.6 |
| 27 | DeepSeek V4.1 Flash | max | 175.6 |
| 28 | DeepSeek V4.1 Flash 1017 | high | 175.4 |
| 29 | Ling 3.1 Flash | 173.8 | |
| 30 | DeepSeek V4.1 Flash 1017 | max | 171.5 |