Skip to main content
Use one stable Fireworks model ID while a model router chooses a model for each new user turn. The selected model handles that turn, including its tool calls, until the next user turn. Family-based router IDs can advance as Fireworks updates the models behind them, without requiring you to change the ID. For example, firerouter/opus makes the Opus family available to the router alongside Fireworks open models. In Claude Code, it may already appear in /model. If it does not, add it:
Restart Claude Code, open /model, and select firerouter/opus. For coding harnesses, start with FireConnect. If your team uses an LLM gateway to access models, follow LLM Gateways. For direct API or SDK calls, see APIs and SDKs.

​
Choose a router

For open-only routing with auto or auto-instant, see Open Models. The routers below add a closed family, pin a closed model, or define custom route members. Use a family name when you want new versions after Fireworks evaluates them. Use a full model ID when you need to pin a version.

​
Choose up to eight route members

List up to eight different models in one route:
Examples:
A custom list defines the models considered during normal routing. The first model is the route’s primary. The router still chooses a model for each new user turn. Every closed model needs a credential for its provider.

​
Pay for the model that serves each turn

You pay for the model that serves each turn:
  • Fireworks models bill to your Fireworks account at their serverless token rates.
  • Closed-model usage bills through that provider’s credential.
  • Cached input bills at the serving model’s cached-input rate.
Check the response model field to see which model served a request. Use Usage and Cost to review a Claude Code session.

​
Watch routing live

In Claude Code, open the harness and a live per-turn meter side by side:
The left pane runs Claude Code. The right pane adds one row for each user turn, including the model or models selected, token buckets, estimated cost, and a prompt preview. Watch it update as the router moves between open and closed models.
Split terminal with Claude Code on the left and a live meter on the right showing GLM 5.3 Flash and Opus 5 model choices, token buckets, cache share, and estimated cost per turn

Claude Code beside the live per-turn model and cost meter

To resume a specific session:
Meter controls and pricing details: Session Cost.

​
Balance quality and savings

Routing preference controls how strongly the router favors predicted quality or lower cost:
Choose from max-intelligence, more-intelligence, balanced, more-savings, and max-savings. Support varies by harness. See Routing Preferences.

​
Connect credentials for closed models

Your Fireworks key is always required. Closed models also need a credential for their provider. Routes that include a Claude family alias or Anthropic model ID require an Anthropic credential. Examples include opus and claude-opus-5-5. Routes that include Astra or another OpenAI model require an OpenAI credential. These rules apply to single-model and custom routes. Account-level Provider Keys are the recommended setup because developers authenticate with only their Fireworks API key. See Harness Compatibility for harness-local alternatives and current FireConnect constraints. See Open Models for aliases, pinned models, image support, and US-only guidance.