Skip to content

Models

What the catalogue holds, how a model is priced, and what happens when one is withdrawn.

The catalogue is the list of models the gateway will route to. Each row carries the four things worth comparing before you pick one: context window, input price, output price, and throughput.

Pricing

Prices are quoted per million tokens and charged per token. Input and output are priced separately, and output is always the dearer of the two.

ModelContextInput / MOutput / MCache
Claude Opus 5200K$1.10$5.50$0.11
Claude Sonnet 5200K$3.00$15.00
GPT-5.6 Luna1.1M$0.20$1.20$0.02
GPT-5.6 Sol1.1M$4.00$20.00$0.40
GPT-5.6 Terra1.1M$2.00$12.00$0.20
DeepSeek V4 Flash1M$0.44$1.32$0.014
DeepSeek V4 Pro1M$1.32$3.96$0.044
Kimi K2.7 Code262K$0.95$4.00$0.19
Kimi K31.0M$3.00$15.00$0.30

A request's cost is prompt_tokens × input + completion_tokens × output, held in micro-dollars and never rounded until it is printed. A tenth of a cent is a real number here, and the table above shows one where a provider charges one.

Cache is the rate for a prompt prefix the provider has already seen. Where a provider does not price cache reads separately the column shows an em dash, which is not the same as free.

The table is the catalogue itself, read live — the same rows the gateway bills from, so there is no second copy of a price here to fall out of date.

Choosing one

  • Cheapest per request — the small model, in front of the others. Intent detection, routing and moderation do not need depth.
  • Long documents — the large context window, where the prompt carries more weight than the completion.
  • Planning and code — the reasoning model, and budget for the extra output tokens it spends thinking.

When a model is withdrawn

Usage is recorded against the model id the gateway actually routed to, not against a catalogue entry. A model that leaves the catalogue keeps its history: the spend still happened, so it still shows up in Usage, under the id it was billed as.