OpenRouter operates on a pass-through model, billing each request at the underlying provider’s list price and collecting a fee when credits are purchased: 5.5% on the Standard plan and 8% on Business, with a $0.80 minimum per purchase. Together AI and Fireworks AI skip that middle layer and sell tokens directly from their own GPU fleets. On the open models both clouds host, standard rates match to the decimal; DeepSeek V4.1 Flash costs $0.30 per million input tokens and $1.20 per million output tokens at either one. OpenRouter is a unified API and routing layer that picks a provider endpoint from a catalog of 400+ models, while Together AI and Fireworks AI own or lease inference capacity, selling tokens, GPU-hours or training time directly. OpenRouter’s documentation prices every model per token, per request or per unit, so the true cost of the routing layer is the credit fee plus whatever the router does with traffic. Together AI and Fireworks AI compete on speed tiers, batch discounts, dedicated GPUs, fine-tuning, and where data gets processed. Meta’s Muse Spark Open landed on Together AI on October 15 and on Fireworks AI on October 20, and OpenRouter rebuilt its default provider routing on October 28.