AI usage-based pricing — the mechanics, the traps, the clauses.

How AI vendors actually price.

AI usage pricing comes in four families. Token-based (OpenAI, Anthropic, Google Gemini) charges per input and output token, with input and output priced separately and at different rates by model. Credit-based (Microsoft Copilot, Adobe Firefly, Salesforce Einstein) abstracts tokens behind a credit denomination, often with credits consumed at varying rates depending on action type. Request-based pricing meters per API call regardless of size. Agent-run pricing meters per autonomous task completion.

In our experience across 340+ engagements, the cost surprises do not come from list pricing — they come from consumption patterns. Output tokens in long-form generation, retrieval-augmented generation (RAG) over large corpora, and agent loops that retry on partial completion each push consumption 3–10x above first-month projections. The contract is signed before the consumption is understood. Modelling that consumption before signature is core to our cloud contract advisory work on AI commitments.

Token economics: input vs output asymmetry

Output tokens are typically 3–5x the price of input tokens. In long-form generation use cases, output dominates cost; in classification or extraction use cases, input dominates. Cost modelling should be use-case-specific, not generic per-token-multiplied-by-volume.

Credit denomination: opacity by design

Credit models abstract underlying token consumption. The advantage to the buyer is simplicity; the disadvantage is reduced visibility into per-action cost. The negotiation move is to demand consumption-tracking telemetry that maps credit consumption back to underlying actions — not just credit-balance totals.

Four AI pricing traps that surface in year one.

1. Pilot-to-production gap. Pilot consumption is non-representative. Production volume is 5–20x pilot. Commit on pilot data and the year-one true-up is brutal.
2. Retry loops in agents. Agentic systems retry on partial completion. A single user request can consume 3–10 underlying credit blocks. Per-request cost in the contract is misleading.
3. Long-context degradation. Vendors price by token; long-context calls are token-dense. RAG-heavy workflows consume materially more than first-pass models suggest.
4. Auto-tier upgrades. Contracts that allow vendor-side model selection or auto-upgrade to newer models can route traffic to more expensive models without buyer consent.

Six clauses that cap usage exposure.

1. Hard usage cap. A contractual ceiling on monthly or annual consumption with overage requiring written buyer consent.
2. Soft cap with notification. Notification at defined thresholds (50%, 75%, 90%) before overage applies.
3. Model lock. The model versions in scope are fixed; auto-upgrade requires buyer consent.
4. Rate card lock. Per-token, per-credit, per-request rates fixed for the contract term — not at vendor discretion.
5. Burst cap and rate limiting. Per-minute or per-hour ceilings to prevent runaway consumption from misconfigured systems.
6. Audit and reconciliation rights. Buyer right to audit vendor's consumption metering and reconcile against buyer's own logs.

Pre-purchase vs PAYG and the right mix.

AI vendors offer pre-purchased capacity (committed token / credit pools, usually multi-year) at a discount against pay-as-you-go. The discount is typically 15–35%. The over-commit risk is real — unused capacity does not roll over indefinitely and is often lost at term end. The under-commit risk is also real — burst consumption hitting list rate when the committed pool is exhausted.

The hybrid commitment structure

The model that consistently performs best in our engagements: a committed pool sized at the 60–70th percentile of projected consumption, with overage at a negotiated rate (not list), and quarterly true-up rather than annual. The structure caps the under-commit penalty while limiting over-commit exposure to a defined window.

Rate card negotiation

Vendor list rates are negotiable at material commitment. Per-token, per-credit, and per-request rates each move 10–30% off list with proper commercial framing — particularly when paired with competitive alternative narratives. The rates apply not only to the committed pool but to overage; the overage discount is often the more valuable cut.