SF·SIX / SOURCE-BACKED RESEARCH
Fastest LLM APIs
Measured output speed (tokens per second) and latency of language-model API endpoints, by provider and service tier, from OpenRouter’s rolling 30-minute measurements.
The fastest LLM API endpoint right now is Relace Apply 3 on Relace at 6,674 tok/s; the lowest latency is Llama 3.1 8B Instruct on Groq at 78 ms. Sample: 1,166 priced endpoints of 318 models with an OpenRouter speed measurement (Previous 30 minutes; reported by OpenRouter).
Highest throughput
| Model | Provider and terms | Throughput | Latency | Input / 1M | Output / 1M |
|---|---|---|---|---|---|
| Relace Apply 3 | Relacefp8 | 6,674 tok/s | 439 ms | $0.85 | $1.25 |
| Morph V3 Fast | Morph | 3,758 tok/s | 399 ms | $0.80 | $1.20 |
| Morph V3 Large | Morphfp8 | 1,959 tok/s | 392 ms | $0.90 | $1.90 |
| gpt-oss-120b | Cerebrasfp16 | 921 tok/s | 178 ms | $0.35 | $0.75 |
| Extract | ScaleDown | 878 tok/s | 942 ms | $0.04 | $0.00 |
| Qwen3.8 27B | Cerebrasfp16 | 600 tok/s | 369 ms | $0.99 | $1.49 |
| gpt-oss-safeguard-20b | Groq | 592 tok/s | 141 ms | $0.075 | $0.30 |
| Beam 1.0 501B A23B | Basetenfp4 | 520 tok/s | 220 ms | $1.30 | $4.30 |
| Beam 1.0 501B A23B | CoreWeavefp4 | 416 tok/s | 280 ms | $1.00 | $3.00 |
| gpt-oss-120b | SambaNova | 415 tok/s | 993 ms | $0.14 | $0.95 |
| gpt-oss-20b | Groq | 414 tok/s | 362 ms | $0.075 | $0.30 |
| GLM 5.2 | DecartFast tierfp4 | 413 tok/s | 423 ms | $2.50 | $9.00 |
| Grok 4.20 Multi-Agent | xAIPriority tier | 388 tok/s | 4.30 s | $2.50 | $5.00 |
| gpt-oss-20b | Amazon Bedrock | 365 tok/s | 385 ms | $0.07 | $0.15 |
| Llama 3.1 8B Instruct | Groq | 361 tok/s | 78 ms | $0.05 | $0.08 |
| GPT-5.6 Luna Pro | OpenAIFast tier | 348 tok/s | 2.08 s | $0.40 | $2.40 |
| Grok 4.20 Multi-Agent | xAIZero data retention | 348 tok/s | 18.9 s | $1.25 | $2.50 |
| gpt-oss-120b | Groq | 344 tok/s | 150 ms | $0.15 | $0.60 |
| MiniMax M2.7 | Groq | 335 tok/s | 268 ms | $0.60 | $1.80 |
| gpt-oss-120b | Crusoebf16 | 331 tok/s | 207 ms | $0.05 | $0.25 |
| Mercury 2 | Inception | 330 tok/s | 596 ms | $0.25 | $0.75 |
| gpt-oss-120b | Nebius AI Studiofp4 | 328 tok/s | 216 ms | $0.15 | $0.60 |
| DeepSeek V4.1 Flash | Fireworks AIUS region | 272 tok/s | 515 ms | $0.45 | $1.80 |
| gpt-oss-20b | Googleus-central1 region | 271 tok/s | 282 ms | $0.07 | $0.25 |
| gpt-oss-120b | Parasailfp4 | 267 tok/s | 346 ms | $0.10 | $0.75 |
Output tokens per second, fastest first. Fast, turbo and priority tiers are labelled; they cost more.
Lowest latency
| Model | Provider and terms | Throughput | Latency | Input / 1M | Output / 1M |
|---|---|---|---|---|---|
| Llama 3.1 8B Instruct | Groq | 361 tok/s | 78 ms | $0.05 | $0.08 |
| Beam 1.0 501B A23B | Together AI | 177 tok/s | 88 ms | $0.40 | $1.20 |
| MiniMax M2.5 | Friendli | 158 tok/s | 91 ms | $0.30 | $1.20 |
| Gemma 4 31B | ModelRunfp4 | 208 tok/s | 98 ms | $0.75 | $1.00 |
| gpt-oss-20b | CoreWeavefp4 | 121 tok/s | 103 ms | $0.03 | $0.13 |
| Llama 3.3 70B Instruct | CoreWeavefp16 | 86 tok/s | 111 ms | $0.71 | $0.71 |
| Mistral Small 4 | Mistral AIUS region | 139 tok/s | 135 ms | $0.165 | $0.66 |
| gpt-oss-safeguard-20b | Groq | 592 tok/s | 141 ms | $0.075 | $0.30 |
| gpt-oss-120b | Groq | 344 tok/s | 150 ms | $0.15 | $0.60 |
| gpt-oss-120b | Cerebrasfp16 | 921 tok/s | 178 ms | $0.35 | $0.75 |
| Qwen3.5-9B | Together AI | 130 tok/s | 181 ms | $0.17 | $0.25 |
| Llama 3 8B Lunaris | DeepInfraTurbo tierfp8 | 92 tok/s | 184 ms | $0.04 | $0.05 |
| Llama 3.2 1B Instruct | Cloudflare | 141 tok/s | 198 ms | $0.027 | $0.201 |
| DeepSeek V3.1 | CoreWeavefp8 | 84 tok/s | 206 ms | $0.55 | $1.65 |
| Seed-2.0-Mini | Seedfp8 | 140 tok/s | 206 ms | $0.10 | $0.40 |
| gpt-oss-120b | Crusoebf16 | 331 tok/s | 207 ms | $0.05 | $0.25 |
| Llama 3.2 3B Instruct | Cloudflare | 236 tok/s | 211 ms | $0.0509 | $0.335 |
| Muse Glimmer 30B | Together AI | 87 tok/s | 215 ms | $0.35 | $1.50 |
| gpt-oss-120b | Nebius AI Studiofp4 | 328 tok/s | 216 ms | $0.15 | $0.60 |
| DeepSeek V4.1 Flash | BasetenUS region | 212 tok/s | 217 ms | $0.30 | $1.20 |
| Beam 1.0 501B A23B | Basetenfp4 | 520 tok/s | 220 ms | $1.30 | $4.30 |
| LFM2.5-24B-A2B | Liquid AIfp8 | 212 tok/s | 223 ms | $0.05 | $0.70 |
| Granite 4.2 8B | CoreWeavebf16 | 92 tok/s | 231 ms | $0.10 | $0.15 |
| Ministral 3 8B 2512 | Mistral AIEU region | 74 tok/s | 242 ms | $0.165 | $0.165 |
| Laguna XS 2.1 | Poolsidefp8 | 140 tok/s | 244 ms | $0.06 | $0.12 |
Latency as reported by OpenRouter for the endpoint, lowest first.
Fastest endpoint of each flagship model
| Model | Measured endpoints | Fastest provider | Throughput | Latency | Standard price (in / out) |
|---|---|---|---|---|---|
| Claude Opus 5.5 | 10 | AnthropicFast tier | 192 tok/s | 1.92 s | $4.00 / $20.00 |
| Claude Sonnet 5.5 | 8 | AzureGlobal endpoint | 119 tok/s | 1.39 s | $2.00 / $10.00 |
| Claude Haiku 5.5 | 9 | GoogleUS region | 189 tok/s | 1.58 s | $0.10 / $0.50 |
| Claude Fable 5.1 | 4 | Amazon Bedrock | 70 tok/s | 7.35 s | $10.00 / $50.00 |
| GPT-6 Astra | 7 | OpenAIUltrafast tier | 157 tok/s | 1.96 s | $10.00 / $50.00 |
| GPT-6 Astra Pro | 4 | OpenAIFlex tier | 83 tok/s | 112 s | $10.00 / $50.00 |
| GPT-6.1 Sol | 8 | OpenAIUltrafast tier | 98 tok/s | 2.10 s | $2.00 / $10.00 |
| GPT-6.1 Sol Pro | 4 | OpenAIFlex tier | 143 tok/s | 100 s | $2.00 / $10.00 |
| GPT-6 Luna | 7 | OpenAIFast tier | 104 tok/s | 1.23 s | $0.10 / $0.50 |
| Gemini 3.8 Flash | 5 | GoogleUS region | 127 tok/s | 1.47 s | $0.75 / $3.75 |
| Gemini 3.5 Flash Lite | 7 | GoogleEU region | 255 tok/s | 1.39 s | $0.30 / $2.50 |
| Grok 4.7 | 5 | xAIZero data retention | 72 tok/s | 1.92 s | $2.00 / $6.00 |
| DeepSeek V4.1 Flash | 35 | Fireworks AIUS region | 272 tok/s | 515 ms | $0.09 / $0.18 |
| DeepSeek V4 Pro 0813 | 21 | Together AI | 110 tok/s | 938 ms | $0.1333 / $0.40 |
| GLM 5.3 | 45 | Fireworks AIFast tier | 197 tok/s | 678 ms | $0.50 / $1.51 |
| GLM 5.3 Flash | 32 | Friendli | 164 tok/s | 712 ms | $0.075 / $0.25 |
| Kimi K3 | 23 | Makorafp4 | 153 tok/s | 812 ms | $0.504 / $8.10 |
| Qwen3.8 Max (0902) | 2 | Alibaba | 33 tok/s | 2.66 s | $2.00 / $6.00 |
| MiniMax M3 | 14 | CoreWeavefp4 | 144 tok/s | 469 ms | $0.23 / $0.96 |
| Muse Wonder | 1 | Meta | 89 tok/s | 2.81 s | $4.00 / $20.00 |
| Muse Spark 1.3 | 1 | Meta | 105 tok/s | 2.73 s | $1.25 / $4.25 |
| MiMo-V2.6-Pro | 4 | DeepInfrafp8 | 34 tok/s | 1.37 s | $0.43 / $0.87 |
| Step 5 Preview | 4 | StepFunfp8 | 163 tok/s | 1.58 s | $1.00 / $2.70 |
| Mistral Large 4 | 3 | Mistral AIEU region | 65 tok/s | 665 ms | $0.68 / $2.09 |
Method and sample
Throughput, latency and uptime are reported by OpenRouter for each provider endpoint over the previous 30 minutes; SF·SIX does not run its own speed tests. Scope: priced endpoints of language models listed on OpenRouter with an SF·SIX model page (1,166 endpoints at the time shown). Speeds move with load and prompt length: treat them as a snapshot, not a guarantee.