One endpoint · Eleven frontier models
Frontier inference,
measured not promised.
GPT, DeepSeek, Kimi, Llama and Grok behind a single OpenAI-compatible API, with up to 10M tokens per minute per model.
Models
Peak tok/s is the fastest median decode speed per stream we measured, at any concurrency level. Lowest TTFT is the quickest median time to first token. Click a column header to sort.
| Model | Peak tok/s | Lowest TTFT | Peak aggregate | Capacity |
|---|
Method: streaming chat completions, about 900–2,000 output tokens per request, lowest available reasoning effort. Runs at 1, 8, 128 and 256 concurrent requests, November 2026. Decode tok/s = output tokens ÷ (last token − first token), so queue time is excluded. Peak aggregate = total output tokens per second across all streams in the best run. Capacity = tokens-per-minute quota per model.
OpenAI-compatible
Point any OpenAI SDK at one base URL and switch models by changing
model.Built for parallelism
Measured at 256 concurrent streams. The fastest models hold 80–100 tok/s per stream with first tokens in about 2s.
High ceilings
Up to 10M tokens and 10K requests per minute per model, and 100K requests per minute on gpt-5.3-codex.