# Fireworks launches FireRouter for Claude Opus workloads

> FireRouter routes coding tasks among three models. Fireworks says internal tests cut session costs 57%, with a small drop in scored accuracy.

Oossa · 2026-10-23 · https://oossa.com/en/fireworks-launches-firerouter-for-claude-opus-workloads

Fireworks AI says FireRouter with Opus is available through its CLI and as a standalone serverless endpoint for any Fireworks account. It routes each user turn among Claude Opus 5.5, GLM 5.3 and GLM 5.3 Flash, weighing task fit, model cost and the value of keeping prompt-cache hits.

In an internal test across its coding workloads, Fireworks reports average cost per session fell from $15.36 to $6.63. Graded-turn accuracy was 78.7%, compared with 80.2% for Opus alone; Fireworks describes that as 98.1% of Opus’s accuracy. The company says the router is supported in Claude Code, Codex and Cursor IDE.

## The facts

- Fireworks reports a 57% cost reduction per session, with a margin of ±19 percentage points.
- Cache hit rate was 94.2% with FireRouter, versus 97.8% with Opus alone.

## Why it matters

Teams using Claude Opus for coding can try Fireworks’ router to reduce session costs, but the reported results are from internal testing and show slightly lower graded-turn accuracy.

## Sources & references

1. [Introducing FireRouter with Opus](https://fireworks.ai/blog/introducing-firerouter-with-opus/) – Fireworks AI
2. [FireRouter with OpusNewRouter1000000 Context](https://fireworks.ai/models/fireworks/firerouter-opus/) – Fireworks AI

Last updated: 2026-10-23