Pricing

Input

Every prompt token the model had to read for the first time. Billed as usage.prompt_tokens.

Cached input

Prompt tokens served from a cached prefix, billed below ordinary input. Applied automatically; cache writes carry no separate fee. Reported asusage.prompt_tokens_details.cached_tokens. Exception, GLM-5.3-Flash: Prompt caching restores retained prefixes and bills only the tokens reported as cached_tokens. In qualification, repeating a 902,579-token prompt reused 902,560 tokens; warm repeat TTFT was 1.819 to 3.972 seconds. Cold prompts still need processing. Residency is bounded and shared; a later request can return cold. Exception, DeepSeek-V4.1-Flash: Prompt caching restores retained prefixes and bills only the tokens reported as cached_tokens. In qualification, a 32.7k-token prompt (32,674-32,758 tokens) reported 32,512 cached tokens on 9 of 9 repeats. Cold prompts still need processing. Residency is bounded and shared; a later request can return cold.

Output

Tokens the model generated, including reasoning tokens when thinking is on. Billed as usage.completion_tokens.

Per request

$0.00. No request fee, no per-key fee, no minimum spend.

Context

The full native window at the rates above. Long prompts bill more tokens, never a higher rate: there is no long-context tier and no separate charge for the window.

Output speed

A measured ceiling on the serving build, input to output: the fastest measured shape, so ordinary requests land below it. First-token figures are published per model, with reasoning off, on the model pages. The measurement protocol.

Priced

The date the current rate was set. Sort by it to see what is newest.

* GLM-5.3-Flash is served with DFlash 2 (inco.ai) as its speculative drafter used by this model, under written permission from its authors.

One prepaid balance and one key cover every model; each request bills at the rate of the model it named. Every live model is served at the same base URL; the model id in the request picks the rate. The quickstart shows the request shape. Every response reports billed token counts split into input, cached input and output, in the standard usage field, in streaming responses too. The same numbers appear in your console.

One prepaid balance and one key cover these too, and they answer the same base URL as every other model. A transcription response reports the billed audio seconds in itsusage field, the same place a chat response reports billed tokens, and the same number appears in your console. Request shapes are in the API documentation.