Fireworks FireRouter with Opus: cache-aware serverless routing

Fireworks published FireRouter with Opus on 2026-09-28—a cache-aware router optimized for the Opus family, available in the Fireworks CLI and, newly, as a standalone serverless router model any Fireworks account can point at “as you would any other model” (Fireworks blog).

This is a Fireworks Nexus coding-route surface—not a multi-provider enterprise AI gateway.

What it routes today

As of the announce, FireRouter with Opus routes among three targets:

• Claude Opus 5.5
• GLM 5.3
• GLM 5.3 Flash

Name them as routing destinations in Fireworks’ set—not as Fireworks-owned models or partnership announces. Fireworks says the model set will change as new models and versions launch (Fireworks blog).

How the cache-aware pass works

Each user turn is scored for how well each model in the set fits the task. The router then estimates cost—including prompt-cache cost and whether losing cache to switch models is worth it—and picks the quality/cost balance. Fireworks frames that prompt-cache awareness as the product’s differentiator; treat “first on the market” comparative lines as vendor marketing, not an independent ranking (Fireworks blog).

Internal A/B figures (Fireworks-attributed)

Fireworks reports more than a month of internal coding-traffic A/B (FireRouter with Opus vs Opus-only control):

• 98.1% of Opus-only accuracy (78.7% vs 80.2% graded turns)
• 57% lower session cost (±19 percentage points), from $15.36 to $6.63 per session
• 94.2% cache hit rate vs 97.8% Opus-alone—described as a small, deliberate tradeoff

All of those numbers are Fireworks internal results, not independent benchmarks and not a public price card (Fireworks blog).

Surfaces and harness color

Beyond the CLI and the standalone serverless endpoint, Fireworks says FireRouter with Opus is available across harnesses including Claude Code, Codex, Cursor IDE, “and more.” Treat that list as availability color, not an exhaustive GA matrix (Fireworks blog).

Pricing (as stated—no invented list rates)

The blog’s session-dollar pair ($15.36 → $6.63) comes from the internal A/B above. The post does not publish a public per-token or list-rate catalog for FireRouter, Opus, or GLM. This write-up invents none (Fireworks blog).

Who should care

Teams already on Fireworks who want a cache-aware coding router they can hit from the CLI or as a serverless endpoint—without standing up a separate multi-provider gateway—should start at the FireRouter with Opus post. Lead with the three current routing targets and the note that the set can change.