AI HOT (Curated Pool)Sep 28, 2026, 00:00 UTC+8

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

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

Fireworks AI packaged its router as a standalone model endpoint that picks between Claude Opus 5.5, GLM 5.3, and GLM 5.3 Flash per turn. In over a month of internal A/B testing on coding tasks, it retained 98.1% of Opus-only accuracy while dropping per-session cost from $15.36 to $6.63—a 57% cut. The cache-aware router deliberately trades roughly 3.6 percentage points of cache hit rate for lower spend, ending at a 94.2% hit rate. It works with Claude Code, Codex, Cursor IDE, and others via two CLI commands.

Why it matters: Fireworks cut Opus 5.5 routing cost by 57% with internal A/B data — real savings for devs coding with Claude. Not p1 because it's a routing-layer optimization, not a model capability leap, and only Fireworks' own numbers, no external validation.

Read the original ↗Export Markdown

Take

This is worth a click because it turns model routing from an internal trick into a product you can actually call. FireRouter with Opus sits between Claude Opus 5.5, GLM 5.3, and GLM 5.3 Flash, picking the cheapest model per turn. The router doesn't just compare quality—it weighs whether switching models is worth losing the prompt cache.

Over a month of internal A/B testing on coding tasks, pure Opus cost $15.36 per session. FireRouter brought that down to $6.63, a 57% drop. Accuracy went from 80.2% to 78.7%, so you keep 98.1% of the quality. Cache hit rate dipped from 97.8% to 94.2%—a deliberate trade to save money.

I'd discount this a bit: the numbers come from Fireworks' own internal coding traffic, not a public benchmark. The post doesn't spell out how they scored accuracy beyond "did the agent finish, was the answer right, did the user correct it next turn." The routing pool is just three models right now, with Opus 5.5 doing the heavy lifting and GLM variants handling cheaper turns. If GLM's future versions wobble, the stability of this setup is an open question.

For teams already on Claude Code or Cursor, two CLI commands get you in. The catch: you only save if enough of your turns don't actually need Opus. If you're deep in hard debugging all day, the router will mostly send work to Opus anyway, and the savings shrink.

Source: AI HOT (Curated Pool) · fireworks.ai
Published: Sep 28, 2026, 00:00 UTC+8 · 2 days ago