Token billing has a place as backend infrastructure. As a customer-facing price, it mostly gets in the way. In its simplest form, passthrough pricing bills the customer for the tokens their usage consumes plus a markup on the underlying cost — which means your costs decide your price.
That split shows up in how companies actually use it. They want token billing to track usage, manage markups and protect unit economics; they don't want those internals shaping their product's price. The guiding principle: the invoice is where you define your value to the customer.
The invoice problem
Token billing got built because it guards against runaway compute costs and is quick to deploy and explain. For model providers, tokens are the product. For most other AI companies, tokens are at least an obvious, countable basis for pricing. But neither tokens nor subscription seats reflect the value the customer receives.
The deeper issue is positioning. Showing a customer which models ran, how many tokens each consumed and what markup was applied frames your product as a commodity markup stacked on an actual commodity. An invoice that itemizes ten models with markups of 10%, 5% and 25% defines your value as the spread between your price and someone else's model costs. As models get cheaper and more interchangeable, customers can dispute the markup or route around you — and competition pushes that markup, and possibly the business, toward zero.
Credits as a presentation layer
Many companies already treat Metronome as a margin-tuning system rather than just a place to deploy pricing models. They make routing decisions upstream and send the resulting token usage in — broken out per customer or rolled up into a single aggregate customer for cost tracking — and get visibility into consumption for cost management and margin tuning.
That handles internal economics, not the invoice. A presentation layer still has to translate raw usage into something tangible. This is where the unified credit model comes in.
A unified credit model works as a credit-based burndown: customers draw on one shared balance as they use different product operations, each task consuming credits at its own rate according to underlying token consumption. One operation might enrich a dataset, another might generate an image. The invoice then shows enrichments or generated images — not the ten models behind them and the markup on each.
The complexity doesn't disappear; it moves behind the scenes. Token-level metering is preserved, so companies can keep routing models upstream and tuning for margin and performance as workloads change or new models arrive. The customer just sees credits tied to the work the product performs instead of tokens tied to model costs.
Inputs, outcomes, outputs
Unified credits are a stepping stone toward output-based pricing, the best way to align AI consumption pricing with value today. Three metrics are in common use.
An input is a cost, like a token.
An outcome is the customer making more money, building more pipeline or reducing churn. Outcomes orient a product around customer value, but they make poor defensible pricing metrics. Outcome-based pricing works only in narrow circumstances — essentially a luxury for the corporate super-rich. You need either monopoly-like power (control over the transaction and its telemetry, plus the market power to make customers accept it) or a very high ticket price, since only very large contracts justify instrumenting a workflow end-to-end and proving the efficiency gain: fewer people, faster work, quicker inventory turns. At that scale you can demonstrate the value every quarter and charge as a function of it. For everyone else, the model is a myth, undone by a telemetry gap: what counts as the outcome, whether it happened, and whether your product caused it.
Charge for a sales agent on the revenue it generates and you invite a dispute over human versus automated labor. Did the agent close the sale, or the sales rep, or marketing — would the customer have bought anyway? The outcome may be valuable without being objectively attributable to the product.
An output is objective and verifiable: an image generated, an email sent. Fin's pricing based on resolved support conversations is often called the poster child for outcome-based pricing; it's really an output. Happy customers are the outcome; a resolved ticket is something you can count. The playbook: define your output metric, call it an outcome, and convince the market you're right.
How pricing models spread
Pricing changes slowly because customers compare you with competitors, and departing from the herd carries a heavy penalty. But once a new model takes hold, the shift can be fast — maladaptive one moment, adaptive and spreading through the category the next.
Three years ago Metronome's team predicted hybrid pricing — seats plus usage — would become a prominent model. Support was built, then went largely unused for about 18 months. Then, as of August 2026, roughly one in six Stripe users past key revenue milestones (and growing) were actively using or rolling out hybrid pricing.
Fin followed the same curve. Three years ago, almost no one in AI customer support talked about outcome-based pricing; within a quarter of Fin's launch, all five leading companies in the category had committed to some version of it. Nothing changed, then everything changed, once the category and its buyers matured enough for demand to cascade.
Agents are the likely trigger for output-based pricing. Subscription and seat pricing have dominated because they're easy for human buyers to understand and compare. An agent can weigh far more variables and price each task closer to its value — a trajectory resembling high-frequency trading, where algorithms operate at a speed and complexity humans can't. As agents evaluate and purchase services, pricing won't be capped by how much complexity a human buyer can absorb, and the work itself becomes easier to measure: every task leaves an audit trail, making activities observable and verifiable without separate monitoring. Expect measurable outputs to become pricing metrics as more work moves through agents — even when they're still marketed as outcomes.



