跳到正文
北京时间
原文
Fireworks AI(网页)·· 2 天前精选AI 评分62

Fireworks AI 推出 FireRouter with Opus,编码任务成本降 57%

Introducing FireRouter with Opus

AI 导读

Fireworks AI 发布 FireRouter with Opus,可在 Claude Opus 5.5、GLM 5.3 和 GLM 5.3 Flash 之间做缓存感知路由,并首次以独立路由模型形式作为 serverless 端点开放。

推荐理由

官方给出了内部 A/B 测试的成本与准确率数据,可帮助团队评估用缓存感知路由替代单一 Opus 的可行性。

正文 · 原文

Fireworks Nexus enables engineering teams to drop leading open models in the harnesses they already use and cut spend in half without sacrificing speed or quality. The solution includes FireRouter, the first cache-aware router on the market, which makes a big difference in speed and cost.

Today, we’re introducing FireRouter with Opus, optimized for the Opus family and now available in both our CLI and, for the first time, as a standalone router model. Any Fireworks account can point to it as a serverless endpoint as you would any other model.

After more than a month of internal A/B testing, FireRouter with Opus executes coding tasks at 98.1% of the accuracy for 57% lower cost versus Opus alone.

What FireRouter is

FireRouter-diagram
Today, FireRouter with Opus routes between Claude Opus 5.5, GLM 5.3, and GLM 5.3 Flash

Each user turn that hits FireRouter gets evaluated on how well each model in the set is suited for the task. It then estimates the cost of each model handling the task, including the cost of the prompt cache and whether losing it to switch to a different model is worth it. And finally, it routes to the model that best balances quality against cost.

At the moment, a user turn will be routed between Claude Opus 5.5, GLM 5.3, and GLM 5.3 Flash. The model set will change as new models and versions launch.

The results

Our internal coding traffic was on sessions randomly assigned to FireRouter with Opus or to an Opus-only control group, across the same workloads and users.

Cost per session fell 57% (±19 ppts), from $15.36 to $6.63.

cost

To score the accuracy, we verified whether the agent finished the work, whether the right answer was reached, and whether the user had to make any corrections on the next turn.

FireRouter with Opus scored 78.7% of graded turns against 80.2% for Opus-only, or 98.1% of its overall accuracy.

Accuracy

Most of our traffic came from internal coding work, where open models could handle many routine turns. FireRouter’s cache-aware routing weighed the savings from switching models against the cost of losing cache hits. The result was a 94.2% cache hit rate with FireRouter and Opus, compared to 97.8% with Opus alone– a small, deliberate tradeoff that substantially reduced overall cost.

Cache

Two lines to try it

FireRouter with Opus is now available across multiple harnesses including Claude Code, Codex, Cursor IDE, and more. Get started with commands below and learn more on FireRouter docs.

bash

Copy

12

fireconnect login

fireconnect claude --model firerouter/opus

Log in to Fireworks to see estimated savings on your team’s Claude Opus spend.

FireRouter with Opus savings chart

来源:Fireworks AI(网页) · fireworks.ai