The company says it routes user turns among Claude Opus 5.5, GLM 5.3, and GLM 5.3 Flash, weighing task suitability, estimated model cost, and the cost of losing prompt-cache hits when switching models.
In more than a month of internal A/B testing on coding traffic, Fireworks reports cost per session fell from $15.36 to $6.63, a 57% reduction (±19 percentage points). Graded-turn accuracy was 78.7% for FireRouter with Opus and 80.2% for Opus-only, which the company describes as 98.1% of the control's accuracy. Cache hit rate was 94.2% versus 97.8%. These are vendor-reported results from internal workloads, not independent evidence of performance on other users' tasks.