Measured
What the APIs did when we called them.
A price says what a token costs. It does not say whether the API answers, how long you wait, or whether the bill matches the price list. Every hour we send each model one small request on every channel we hold a key for, and write down what happened.
Calls answered4028 / 403199.9%
Charges matching the recorded list price1547 / 2330requests where the provider reported the charge
Models and channels measured41one small request each, every hour
By model and channel
| Model | Channel | Answered | First tokenmedian / 9 in 10 within | Speedtokens per second | Charged vs list price |
|---|---|---|---|---|---|
| Claude Fable 5.1 | Claude API | 100%106 / 106 | 0.64 s1.8 s | 125 | — |
| Claude Fable 5.1 | OpenRouterGoogle, Azure | 100%107 / 107 | 1.9 s2.5 s | 135 | as listed |
| Claude Fable 5.2 | Claude API | 100%106 / 106 | 0.64 s0.84 s | 163 | — |
| Claude Haiku 5.5 | Claude API | 100%106 / 106 | 0.53 s0.88 s | 379 | — |
| Claude Haiku 5.5 | OpenRouterClaude Platform on AWS, Google | 100%107 / 107 | 0.56 s0.84 s | 376 | as listed |
| Claude Opus 5.5 | Claude API | 100%106 / 106 | 0.69 s1.2 s | 161 | — |
| Claude Opus 5.5 | OpenRouterClaude Platform on AWS, Amazon Bedrock | 100%107 / 107 | 1.3 s1.7 s | 162 | as listed |
| Claude Sonnet 5.5 | Claude API | 100%106 / 106 | 0.62 s0.83 s | 228 | — |
| Claude Sonnet 5.5 | OpenRouterClaude Platform on AWS, Amazon Bedrock | 100%107 / 107 | 0.57 s0.68 s | 242 | as listed |
| DeepSeek V4 Pro | DeepSeek API | 100%64 / 64 | 1.4 s1.7 s | 474 | — |
| DeepSeek V4 Pro | OpenRouterRelace, Baidu | 100%107 / 107 | 1.8 s3.7 s | 499 | −13% |
| DeepSeek V4.1 Flash | DeepSeek API | 100%64 / 64 | 1.0 s1.2 s | — | — |
| DeepSeek V4.1 Flash | OpenRouterWafer, InferenceNet | 100%107 / 107 | 0.55 s1.1 s | 282 | −48% |
| Gemini 3.5 Flash-Lite | Gemini Developer API | 100%106 / 106 | 0.65 s0.92 s | 570 | — |
| Gemini 3.5 Flash-Lite | OpenRouterGoogle | 100%107 / 107 | 0.60 s0.76 s | 495 | as listed |
| Gemini 3.8 Flash | Gemini Developer API | 100%106 / 106 | 2.6 s3.5 s | 282 | — |
| Gemini 3.8 Flash | OpenRouterGoogle, Google AI Studio | 100%107 / 107 | 2.6 s4.0 s | 301 | −9.1% |
| Gemini 4 Argon | Gemini Developer API | 99.1%105 / 106 | 4.2 s9.5 s | 125 | — |
| Gemini 4 Argon | OpenRouterGoogle, Google AI Studio | 100%107 / 107 | 3.1 s8.1 s | 87 | as listed |
| GLM-5.3 | OpenRouterWafer, Decart | 100%107 / 107 | 0.71 s1.5 s | 253 | as listed |
| GLM-5.3 | Z.AI API | 100%64 / 64 | 1.8 s2.7 s | 282 | — |
| GLM-5.3-Flash | OpenRouterInferenceNet, Wafer | 100%107 / 107 | 0.88 s1.9 s | 78 | +276% |
| GLM-5.3-Flash | Z.AI API | 100%64 / 64 | 1.6 s2.4 s | 168 | — |
| GPT-6 Astra | OpenAI API | 100%106 / 106 | 2.2 s3.1 s | 152 | — |
| GPT-6 Astra | OpenRouterOpenAI | 100%107 / 107 | 1.5 s2.0 s | 148 | as listed |
| GPT-6 Luna | OpenAI API | 100%106 / 106 | 0.95 s1.5 s | 166 | — |
| GPT-6 Luna | OpenRouterOpenAI | 100%107 / 107 | 1.2 s2.1 s | 225 | as listed |
| GPT-6.1 Sol | OpenAI API | 100%106 / 106 | 1.8 s2.2 s | 175 | — |
| GPT-6.1 Sol | OpenRouterOpenAI | 100%107 / 107 | 1.1 s1.5 s | 160 | as listed |
| GPT-6.1 Terra | OpenAI API | 100%106 / 106 | 1.1 s1.5 s | 91 | — |
| GPT-6.1 Terra | OpenRouterOpenAI | 100%107 / 107 | 0.95 s1.4 s | 95 | as listed |
| Kimi K3 | Kimi API | 100%64 / 64 | 3.7 s6.8 s | 166 | — |
| Kimi K3.1 Pro | Kimi API | 98.4%63 / 64 | 2.8 s3.7 s | 88 | — |
| Kimi K3.1 Pro | OpenRouterWafer, InferenceNet | 100%107 / 107 | 1.1 s2.5 s | 285 | −6.5% |
| Llama 4 Maverick | OpenRouterParasail | 99.1%106 / 107 | 0.34 s0.49 s | 90 | as listed |
| MiniMax M3 | MiniMax API | 100%64 / 64 | 0.65 s1.0 s | 164 | — |
| MiniMax M3 | OpenRouterCoreWeave, Parasail | 100%107 / 107 | 0.72 s2.2 s | 200 | −21% |
| MiniMax M3.1 Flash | MiniMax API | 100%64 / 64 | 1.2 s3.0 s | 155 | — |
| MiniMax M3.1 Flash | OpenRouterTogether, Nebius | 100%107 / 107 | 0.55 s0.97 s | 666 | +58% |
| Qwen3.8 Flash | OpenRouterAlibaba | 100%107 / 107 | 1.3 s2.1 s | 170 | as listed |
| Qwen3.8 Max | OpenRouterAlibaba | 100%107 / 107 | 2.5 s3.4 s | 234 | as listed |
How to read it
- Answered
- The share of our hourly calls that returned an answer. Failed calls count against this; calls refused because of our own account (no credit, or a key without access) are excluded.
- First token
- Time from sending the request to the first visible text, measured from our Los Angeles server. Models that think before answering take longer by design.
- Speed
- Tokens per second while writing, from one longer request a day. A dash means no speed run has been made yet.
- Charged vs list price
- Where the provider reports the charge for a request, we compare that amount with the list price we had on record that day. A marketplace routes to different providers at different prices, so its charge can be above or below its listed price. A dash means the provider does not report per-request charges.
- What this is not
- This does not test answer quality or performance under load: one small request from one location says nothing about behaviour under heavy use or from other regions.