Token-count billing is the better billable unit for monetizing B2B AI workflows, because token count moves with your inference cost and execution time doesn't. Execution time earns its place only where the workflow spends most of its wall clock outside the model. Flexprice meters both units on one event stream.

Key Takeaways

• Token count is the only unit that moves with what an LLM workflow costs you, which makes it the default for margin protection.

• Execution time fits workflows dominated by non-LLM work: browser automation, long tool calls, sandboxed code execution.

• Neither unit is predictable for the buyer, so B2B contracts wrap both in an included allowance with overage above it.

• Metering wall-clock time is harder than counting tokens, because queue time, retries and streaming all need defining in the contract.

• Flexprice counts tokens and duration as separate metrics on one event stream, with an event debugger that makes a disputed line traceable.

Which billable unit should I choose for an AI workflow?

Pick the unit that tracks your marginal cost. If the model is the expensive part of the workflow, tokens track it. If the model is a small slice and the workflow spends minutes in tool calls or a sandbox, seconds track it better.

• Token count when the workflow is prompt, retrieve, generate. Cost scales with context length and output length. Our AI pricing metrics guide covers how to choose between candidates.

• Execution time when the workflow drives a browser, runs code, or waits on third-party APIs you pay for by duration.

• Per-run when both are stable enough for one number to cover the spread, which is rare once real inputs arrive.

• Both when the workflow mixes them, priced as tokens plus a compute-minute line.

How predictable is each pricing model for the buyer?

Neither is predictable on its own. Token count varies with input size the customer controls, while execution time varies with retries and third-party latency they don't, which is the harder sell in procurement.

• Tokens are auditable: the customer counts their own prompts and checks your invoice.

• Seconds are opaque: a slow vendor API inflates the bill and nobody can see why.

• An allowance plus overage gives either unit the forecastable number procurement needs.

What are the margin implications of each billing unit?

Token count protects margin because it's the same unit your model provider charges you in. Execution time exposes you, since a workflow that stalls on a retry burns your compute and your customer's patience while billing the same as a clean run.