Skip to content

Standard and lite modes

Some models on Keln have two modes, standard and lite. Models & pricing shows a lite price for each model that has one.

  • Standard is the default, and every model has it. Keln routes the request as described in Routing & failover.
  • Lite runs the same model at a lower price. Keln sends a lite request only to providers that serve the model at the lite price. If none of them can take the request, Keln serves it in standard mode at the standard price.

Both modes run the same model, with the same reasoning levels, tools, caching and zero data retention. Keln checks lite providers with the same quality verification as every other provider.

Turning lite on

Set service_tier to "lite". The field works the same way on all three APIs:

{ "model": "openai/gpt-oss-120b", "service_tier": "lite", "messages": [...] }
{ "model": "openai/gpt-oss-120b", "service_tier": "lite", "input": "..." }
{ "model": "openai/gpt-oss-120b", "service_tier": "lite", "max_tokens": 1024, "messages": [...] }

To make lite the default for a key, open API keys in the portal and set the key's Mode to Lite. A request's service_tier overrides the key: "standard" runs a request in standard mode on a lite key. A client that sends no service_tier follows the key's mode.

In the Playground, the Mode switch sets standard or lite for each message.

Accepted values

service_tier Mode
"lite" lite
"standard" standard
"default" (OpenAI's name) or "standard_only" (Anthropic's) standard
"auto", or no field the key's mode: standard unless the key is set to lite

Keln returns 400 invalid_service_tier for any other value.

Speed targets

Lite meets the same speed targets as standard: at least 50 tokens per second and a first token within 2 seconds. Keln sends a lite request only to providers that meet both targets right now. A lite request can still run slower than the fastest standard provider.

Which mode served a request

Every response says which mode served it:

  • the x-keln-mode header: lite or standard;
  • service_tier in the response body, and in the last chunk of a stream;
  • usage.cost, at that mode's price.

When you asked for lite and Keln served the request in standard mode, the x-keln-lite-fallback header says why:

Reason Meaning
no_lite_price the model has no lite price
no_lite_pair no provider serves the model at the lite price right now
slo no lite provider meets the speed targets right now
unavailable the lite providers could not take this request

Treat any other value the same way: Keln served the request in standard mode and billed the standard price.

Billing

Keln bills each request at the price of the mode that served it, and usage.cost shows that amount. Keln sets the mode when the first token arrives. If a provider fails mid-stream and Keln resumes the stream on another provider, the request keeps its price.

The Usage page counts your requests served in lite mode and shows how much lite saved you.