Routing Preferences

Routing preference controls how strongly a firerouter request favors its primary model or lower-cost models. Set it per HTTP request with x-routing-preference, or configure it for a supported harness with FireConnect.

Preference levels

Send an integer from 1 (strongest preference for the primary model) to 5 (strongest preference for lower-cost models):

Value	Name	Behavior
`1`	`max-intelligence`	Keep the turn on the primary model
`2`	`more-intelligence`	Lean toward the primary, compared with balanced
`3`	`balanced`	Default when the header is omitted
`4`	`more-savings`	Lean toward lower cost, compared with balanced
`5`	`max-savings`	Prefer the lower-cost models in the route

HTTP header

The HTTP header accepts only the integers 1 to 5. Names such as balanced are FireConnect CLI values, not HTTP header values. Set the header on an API request:

curl https://api.fireworks.ai/inference/v1/chat/completions \
  -H "Authorization: Bearer $FIREWORKS_API_KEY" \
  -H "x-anthropic-api-key: $ANTHROPIC_API_KEY" \
  -H "x-routing-preference: 4" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "firerouter",
    "messages": [{"role": "user", "content": "Add a docstring to this function."}]
  }'

If the header is missing, invalid, or out of range, the router uses balanced behavior. It does not add a literal x-routing-preference: 3 header.

FireConnect

When you connect a harness to firerouter or a supported firerouter/... ID, pass --routing-preference with the connect command:

fireconnect claude --model firerouter --routing-preference 4
fireconnect opencode --model firerouter --routing-preference 2
fireconnect pi --model firerouter --routing-preference 5

Harness	Supported model IDs
Claude Code	Bare `firerouter` only
OpenCode, Pi, VS Code	`firerouter` or any ID beginning with `firerouter/`
Codex, Cursor IDE, Copilot, DeepSeek Harness	Not supported

FireConnect accepts values 1 to 5 or the corresponding level names. Claude Code returns an error if you combine --routing-preference with a compound ID such as firerouter/opus. Cursor CLI is not supported.

When to adjust

High-volume workloads with many simple requests: try 4 or 5 so summaries, formatting, and straightforward Q&A are more likely to use a lower-cost model.
Tasks where quality is critical (security review, complex reasoning, nuanced writing): try 1 or 2 so the router favors its primary model.
Evaluating routing: start at 3 (balanced), then move one step at a time and compare cost and output quality.

FireRouter: firerouter, firerouter/opus, and firerouter/astra
APIs and SDKs: HTTP headers and Provider Keys
FireConnect: enable a router in a coding harness

Nexus

Connect

Choose models

Operate

Reference

Preference levels

HTTP header

FireConnect

When to adjust

​Preference levels

​HTTP header

​FireConnect

​When to adjust

​Related

Preference levels

HTTP header

FireConnect

When to adjust

Related