Skip to main content
FireRouter is a managed routing service for any LLM workload. Request the FireRouter model through the Fireworks inference API, and FireRouter scores each turn to decide whether it can be served by a cheaper Fireworks open model or should pass through to a closed-source model. The result is lower cost on simpler requests without giving up closed-source quality on harder ones.
FireRouter is in research preview. APIs, routing behavior, and the models in the routing pair may change. This documentation is updated to reflect the current configuration.

​
When to use FireRouter

Use FireRouter when you want automatic cost optimization without picking a different model per request:
  • You want closed-source quality (for example Claude Opus) on hard prompts but do not need it on every call.
  • Many of your requests are straightforward (summaries, formatting, simple Q&A) and can be served by a Fireworks open model.
  • You are fine sending both a Fireworks API key and a provider API key on each request.

​
How it works

FireRouter scores each request independently and picks one of two paths: FireRouter uses a bring-your-own-key (BYOK) model:
  • Your Fireworks API key authenticates to FireRouter and pays for redirected calls.
  • Your provider API key pays for pass-through calls to closed-source models.
FireRouter never stores your provider keys server-side. You send them on each request.

​
FireConnect

The easiest way to use FireRouter in coding harnesses is FireConnect. As of FireConnect v0.9.0, select FireRouter like any other model:
See FireConnect Models for how firerouter fits next to other short IDs, and Harness support for which harnesses support FireRouter. Per-harness details (Claude slot flags, Codex Anthropic env, routing preference) live on each harness page. Upgrade FireConnect before enabling FireRouter on an older install. See Upgrade FireConnect. Pass --anthropic-api-key sk-ant-... on on, or store a key once with fireconnect configure --anthropic-api-key sk-ant-.... Without an Anthropic key, FireRouter still redirects to open models but cannot pass through to Claude Opus 4.8. See Authentication. Cursor and Deep Agents need workspace BYOK for FireRouter (no local Anthropic key).

​
Endpoint

FireRouter is available through the Fireworks inference API:
Common API paths:

​
Model ID

FireRouter is a first-party model on the Fireworks provider. Use firerouter in the model field:
These longer forms are also accepted:
For LiteLLM, use the full router path in litellm_params.model. FireRouter routes each request between a configured closed-source model and a Fireworks open model. You do not pick the target model per request; FireRouter decides based on request complexity.

​
Current routing pair

The FireRouter model currently routes between: These models are subject to change as FireRouter is updated. This page reflects the current configuration. Because the pass-through target is Claude Opus 4.8, you must supply an Anthropic API key. See Authentication.

​
Client integrations

​
What FireRouter is not

  • Not a deployment router. FireRouter is different from deployment routers, which load-balance traffic across your own Fireworks deployments.
  • Not request-sticky. Each request is scored independently.

​
Next steps

FireConnect

Enable FireRouter with one command

Quickstart

Make your first API call

Claude Code (manual)

Edit settings.json for FireRouter in Claude Code

Authentication

BYOK headers and API key requirements

Routing preferences

Tune the quality vs. savings dial

LiteLLM

Add FireRouter to LiteLLM Proxy