Fireworks AI launches FireRouter with Opus, cutting coding costs 57%

What FireRouter with Opus does

FireRouter with Opus is optimized for the Opus family and is now available in the Fireworks CLI and, for the first time, as a standalone router model. According to Fireworks, any Fireworks account can point to it as a serverless endpoint, as with any other model.

The router is part of Fireworks Nexus, which the company says lets engineering teams drop leading open models into the harnesses they already use and cut spend in half without sacrificing speed or quality. Fireworks describes FireRouter as the first cache-aware router on the market.

Currently, a user turn is routed between Claude Opus 5.5, GLM 5.3 and GLM 5.3 Flash. Fireworks says the model set will change as new models and versions launch.

How the routing works

Each user turn that reaches FireRouter is evaluated for how well each model in the set suits the task. The router then estimates what it would cost for each model to handle the turn. That estimate includes the cost of the prompt cache, and whether losing the cache by switching to a different model is worth it.

Finally, it routes the turn to the model that best balances quality against cost. The cache awareness is the central idea: a cheaper model is only chosen when the savings outweigh the price of abandoning cached context.

Test results reported by Fireworks

Fireworks ran more than a month of internal A/B testing. Internal coding sessions were randomly assigned either to FireRouter with Opus or to an Opus-only control group, across the same workloads and users.

• Cost: cost per session fell 57% (±19 ppts), from $15.36 to $6.63.
• Accuracy: FireRouter with Opus scored 78.7% of graded turns against 80.2% for Opus-only, or 98.1% of its overall accuracy.
• Cache hits: the hit rate was 94.2% with FireRouter and Opus, compared to 97.8% with Opus alone.

To score accuracy, Fireworks checked whether the agent finished the work, whether the right answer was reached, and whether the user had to make corrections on the next turn. The company calls the lower cache hit rate a small, deliberate tradeoff that substantially reduced overall cost. It also notes that most of its traffic came from internal coding work, where open models could handle many routine turns.

Availability and what it means for teams

FireRouter with Opus is available across multiple harnesses, including Claude Code, Codex and Cursor IDE. To try it with Claude Code, users log in with fireconnect login and then run fireconnect claude --model firerouter/opus. Fireworks also says that logging in to its platform shows estimated savings on a team's Claude Opus spend.

For teams that rely heavily on Opus for coding agents, the reported results suggest that routine turns can be handled by cheaper open models with a small accuracy loss. The figures come from Fireworks' own internal traffic, so results on other workloads may differ.