Speed comparison
Output throughput in tok/s and time to first token in seconds. Throughput sets how long a long answer takes; latency sets how the conversation feels. The two often pull in opposite directions.
Speed × intelligence
Top-right is fast and smart — and almost nothing sits there. Speed and intelligence trade off inside the architecture: MoE models run fast because few parameters activate, and slow because their reasoning chains are long.
Throughput ranking
| # | Model | tok/s | Vs. fastest | TTFT s | Intelligence | Output $ |
|---|---|---|---|---|---|---|
| 1 | Claude Haiku 5.5Anthropic | 235 | Fastest | 0.6 | 36.2 | $0.50 |
| 2 | gpt-oss-120b Open weightsOpenAI | 220 | 94% | 0.7 | — | $0.20 |
| 3 | Gemma 4 31B Open weightsGoogle | 200 | 85% | 0.5 | 19.0 | $0.40 |
| 4 | Qwen3.8 27B Open weightsAlibaba Qwen | 180 | 77% | 0.8 | 33.7 | $0.40 |
| 5 | Gemini 3.8 FlashGoogle | 160 | 68% | 1.1 | — | $3.75 |
| 6 | DeepSeek V4.1 Flash Open weightsDeepSeek | 148 | 63% | 2.3 | 32.5 | $1.20 |
| 7 | Muse WonderMeta | 143 | 61% | 1.4 | 43.3 | $10.00 |
| 8 | Claude Sonnet 5.5Anthropic | 135 | 57% | 2.1 | 54.2 | $10.00 |
| 9 | GPT-6 LunaOpenAI | 130 | 55% | 1.2 | — | $0.50 |
| 10 | MiniMax M3 Open weightsMiniMax | 126 | 54% | 1.3 | 29.2 | $1.00 |
| 11 | GLM-5.3 Flash Open weightsZ.ai | 120 | 51% | 1.5 | 41.8 | $0.80 |
| 12 | MiniMax M3.1 FlashMiniMax | 118 | 50% | 1.8 | 29.0 | $0.60 |
| 13 | Grok 4.7xAI | 105 | 45% | 1.8 | 35.0 | $6.00 |
| 14 | Claude Opus 5.5Anthropic | 96 | 41% | 2.4 | 58.4 | $20.00 |
| 15 | Step 5 PreviewStepFun | 95 | 40% | 1.6 | — | $2.00 |
| 16 | GPT-6.1 TerraOpenAI | 94 | 40% | 1.9 | 51.0 | $5.00 |
| 17 | Qwen3.8 MaxAlibaba Qwen | 88 | 37% | 2.0 | — | $8.00 |
| 18 | Gemini 4 CarbonGoogle | 87 | 37% | 2.7 | 59.4 | $10.00 |
| 19 | GPT-6 AstraOpenAI | 81 | 34% | 2.8 | 56.9 | $50.00 |
| 20 | MiMo-V2.6 Pro Open weightsXiaomi | 81 | 34% | 2.9 | 30.0 | $1.40 |
| 21 | Gemini 4 ArgonGoogle | 80 | 34% | 2.6 | 50.4 | $10.00 |
| 22 | GPT-6.1 SolOpenAI | 77 | 33% | 2.5 | 54.5 | $10.00 |
| 23 | GLM-5.3 Open weightsZ.ai | 77 | 33% | 2.2 | 32.0 | $2.20 |
| 24 | DeepSeek V4 Pro Open weightsDeepSeek | 72 | 31% | 2.9 | 23.0 | $3.96 |
| 25 | Nemotron 3 Ultra Open weightsNVIDIA | 70 | 30% | 3.1 | — | $1.60 |
| 26 | Claude Fable 5.1Anthropic | 57 | 24% | 6.2 | 49.7 | $50.00 |
| 27 | Kimi K3 Open weightsMoonshot AI | 30 | 13% | 4.2 | 43.6 | $15.00 |
| 28 | Kimi K3.1 Pro Open weightsMoonshot AI | 28 | 12% | 4.5 | 46.0 | $10.00 |
How to read these numbers
Chat: watch latency
Past 3 seconds to first token, users feel the lag. Reasoning models can take 100+ seconds to their first token including thinking — fine for batch jobs, unusable for live interaction.
Long answers: watch throughput
Generating 4,000 characters takes two and a half minutes at 28 tok/s and 17 seconds at 235 tok/s. For report writing and batch generation that gap converts directly into human-hours.
Fast is not the same as cheap
Haiku 5.5 is fast and cheap, but it cannot run a long-horizon agent. Its speed only helps when you do not need it to think for a long time.