What actually drives your Azure AI Foundry bill, in plain terms

Azure AI Foundry pricing is consumption-based with no flat license fee, which sounds simple until a team ships its first production feature and the bill shows up wildly different from what anyone estimated. The core unit is the token — every model call is billed separately for input and output tokens, and the rate depends heavily on which model handled the request.

The spread between models is bigger than most teams expect going in. GPT-4o runs around $2.50 per million input tokens and $10 per million output tokens at current pay-as-you-go rates, while GPT-4o-mini runs roughly $0.15 and $0.60 for the same — about 94% cheaper on input alone. For most organizations, inference is 50-80% of total Foundry spend, which means model selection is often the single highest-leverage cost decision on the table, ahead of infrastructure choices.

The deployment model matters almost as much as the model itself. Pay-as-you-go suits unpredictable or low-volume workloads; commitment-based pricing can cut costs up to 70% for steady, predictable usage. One consulting group's case data showed a client cutting monthly inference costs from $12,000 to $4,800 simply by routing simple queries to a cheaper model and reserving the expensive model for genuinely complex requests — the same “right-sizing” instinct that already applies to VM reservations.

The practical starting point for a team new to this: before committing to a pricing tier, build a routing layer — even a simple one — that sends easy queries to the cheapest model that can handle them, and reserve the expensive model for the requests that actually need it. It's usually the fastest way to see a meaningful drop in the bill without changing what the product does.