FireRouter with Opus Cuts Opus Costs 57% via Cache-Aware Routing

Fireworks AI introduced FireRouter with Opus, a cache-aware router optimized for the Opus model family. The router is available as a CLI tool and as a standalone router model, and any Fireworks account can use it as a serverless endpoint like any other model. The company announced the solution on September 28, 2026.

Routing and Cost Savings

FireRouter routes each user turn among Claude Opus 5.5, GLM 5.3, and GLM 5.3 Flash. The model set will evolve as new models and versions are released. In internal A/B testing with coding traffic randomly assigned to FireRouter or an Opus-only control, FireRouter showed 57% lower cost versus using Opus alone. Cost per session dropped from $15.36 to $6.63, with a margin of ±19 percentage points. Accuracy during testing was 98.1% of Opus-only accuracy, measured as 78.7% versus 80.2% graded turns.

The router evaluates model suitability, estimated cost, and cache impact before routing, balancing quality against cost for each turn. FireRouter's routing decision includes the cost of losing a prompt cache when switching models. This cache-aware approach deliberately trades a small drop in cache hits for larger cost savings. In the test, cache hit rate with FireRouter was 94.2%, compared with 97.8% for Opus alone.

Integration and Access

FireRouter with Opus is integrated into Claude Code, Codex, Cursor IDE, and other harnesses. To try it, users run the fireconnect login command, then fireconnect claude with the model flag set to firerouter/opus. The solution can also be pointed to as a serverless endpoint via the Fireworks platform.

Fireworks AI markets FireRouter as the first cache-aware router on the market. The company also highlighted Fireworks Nexus, which lets teams drop open models into existing harnesses and halve spend.

For inference engineers, the key takeaway is that routing decisions now factor in cache economics, not just model quality. The trade-off between a 3.6 percentage point drop in cache hits and a 57% cost reduction is a concrete example of how cache-aware routing can shift the cost-quality balance in production workloads.