On Two Platforms, Token Prices Fell and Spend Did Not: AI Cost Control Moves to Entitlement

On OpenRouter, the average price paid for a million tokens fell from $0.77 to $0.14. Separately, the platform reports that its weekly token count went from about half a trillion in January 2025 to 164 trillion by late September 2026. Those two numbers come from a guest post on OnlyCFO by Peter Walker, Head of Insights at OpenRouter, and they describe OpenRouter’s own customer mix, which is one platform and one population of buyers.

Max Woolf, working from Vercel AI Gateway data, reports the same shape on a second platform. Cost per token on the gateway fell “to 43.3% of the April 2026 cost. However, aggregate spend has been increasing.”

Two platforms, two datasets, one direction. On both, the platform reports a cheaper unit alongside growing usage or total spend.

Our reading of that pattern is narrow and practical. When the price of a unit drops and the number of units climbs, any budget written as “we expect to pay roughly this much per token” goes stale quickly. The forecast was correct about price and wrong about everything that price unlocked. A finance team that keeps tuning the price assumption is adjusting the variable that is already moving in its favor.

The control that holds up under this pattern is entitlement: who may spend, how much per week, and who approves more.

Why price-based budgets drift

A per-token budget is a bet on two numbers at once: unit price and volume. Vendors publish the first. Nobody publishes the second for your organization, because it depends on how many people adopt the tool, how many agents run unattended, and how many retries each task takes.

The OpenRouter post gives a clean example of how retries inflate the real unit. “A request that costs $1.00 and succeeds 80% of the time costs $1.25 per successful answer.” The price on the invoice is $1.00. The price of the outcome is a quarter higher. A budget built on the list price misses that difference before a single new user shows up.

We have written before about agents that approve their own overage, about showing people what a prompt costs, and about cost per completed task as the CFO’s number. This piece builds on a different question. Assume the meter exists and the per-task math is understood. What actually stops the spend?

The two vendors below announced answers in October, and both answers are about the person, with the token price left out of the policy entirely.

Harness: a weekly cap with an approval path

Harness introduced AI Budgets with a mechanism that is easy to describe. “At 90% of budget, spend moves into a WARN tier.” Once weekly spend crosses the budget, “access pauses until the next weekly reset or until someone approves a request for more.”

Those two sentences contain several design choices, and each answers a weakness of price-based budgeting.

The first is the cadence. Harness argues that “a monthly budget lets someone burn the whole forecast in week one.” A weekly window bounds the damage of a bad week to a week. In Harness’s words, “a weekly reset keeps the feedback loop tight and makes a breach cheap to recover from.” That is four or five resets per month instead of one.

The second is the warning tier. At 90%, Harness notifies the person and restricts nothing yet. That turns the cap from a surprise into a conversation.

The third is the approval path. Access pauses, and someone can approve more. Harness routes the request by size: “Small increases clear quickly, and larger ones go to the person with the authority to approve them.” Our recommendation is that this person be a named owner.

Harness also reports a number from running the policy on itself: “Average weekly AI spend per developer fell from $184 to $115, about 37% lower.” Treat that figure with care. It is one company testing its own product on its own bill, reported by the vendor. It shows the mechanism can bite inside one organization. It says nothing yet about yours.

Figma: a credit limit provisioned through SCIM

Figma took the same idea and moved it into identity. According to UX News, Enterprise admins can set per-user access to paid AI credits as a monthly limit, full access, or none. The limits cover purchased credits only, and “a personal limit does not reserve credits: shared pools can run out first.”

The more interesting part is where the limit lives. Figma’s help page on managing AI credit limits with SCIM documents an attribute called aiCreditLimit. It takes a number of paid credits per month, or blocked, inherit, or unlimited. When SCIM manages the limit, “their identity provider becomes the source of truth,” and admins can view the limit in Figma but cannot change it there.

That puts AI spend entitlement in the same provisioning pipeline that already decides who gets which application and which group. Our reading: if the identity provider maps the attribute to groups, a new hire could get a credit limit along with the account, and a role change could carry the limit with it. The finance question “how much can this person spend on AI” would then be answered by the same system that answers “does this person get an account.”

Harness and Figma differ on cadence (weekly against monthly), and Figma also lets the identity provider own the limit through SCIM. Harness can scope a budget to a team, a developer, a model, or a project. Figma sets limits per user, with defaults by billing group or organization. Neither policy is expressed in tokens.

What entitlement controls look like in practice

Pulling the two designs together gives a short list of properties worth demanding from any AI tool you buy or any internal gateway you build.

A per-person or per-group ceiling. The unit of control is the human or the team, because that is the unit a manager can reason about. A token is the vendor’s unit.

A short reset window. Harness’s argument about monthly budgets applies to anything with a long window. The shorter the window, the smaller the worst case.

A warning before the stop. A cap that triggers without notice can push people to route around it. A warning tier gives them time to ask.

A named approver for more. When the cap is hit, the path forward is a request to a person with authority, recorded somewhere. That record becomes your demand signal: which teams keep asking, and for what.

Provisioning from identity. Figma’s SCIM attribute shows the direction. If entitlement lives in the identity provider, onboarding, role changes, and offboarding update spend limits automatically. If it lives in a spreadsheet, it drifts the same way the price-based budget did.

None of this requires predicting next quarter’s token price. That is the practical advantage. The price can keep falling, the volume can keep climbing, and the entitlement still holds because it never depended on either.

The objection worth taking seriously

A weekly cap can slow down good work. A developer in the middle of a migration hits the ceiling on Thursday and waits for an approver who is in meetings. That cost is real.

Both designs leave a way to raise the limit. Harness pauses access until a reset or an approval. In Figma, inherit applies the billing group’s default, else the organization’s, and an explicit limit or unlimited overrides it for that user. For a SCIM-managed user, the request flow does not submit a request in Figma, so the change has to happen in the identity provider. The cap is where the conversation starts. If approvals become routine for a team, raise that team’s ceiling and keep the cap for everyone else.

The other objection is that spend is a lagging proxy for value. That is true, and it is why the cost per completed task remains the number to report upward. Entitlement controls answer a different question: how much exposure the organization accepts while it is still learning what that per-task number is.

Do this now

Pick the AI tool with the largest line on last month’s invoice. Answer four questions about it.

Who can spend on it today, by name or by group? What happens when one of those people spends far above the average in a week? Who gets told, and how fast? Who can approve more, and where is that approval recorded?

If the answers are “everyone,” “nothing,” “nobody until month end,” and “nobody,” you have a price-based budget and no entitlement. Start with a weekly per-person ceiling, a warning at 90%, and one named approver per team. If the tool supports provisioning limits through your identity provider, map the ceiling to existing groups instead of maintaining a separate list.

Then watch the approval requests for a month. They are likely to say more about where AI is creating value than a token forecast would.

This analysis synthesizes The Current State of AI Tokenomics | What CFOs Should Know (OnlyCFO, guest post by OpenRouter, October 2026), OpenAI and Anthropic cut their LLM prices by up to 80% within a few months. Why? (Max Woolf, October 2026), Introducing AI Budgets: A Policy You Can Watch Work (Harness, October 2026), Figma Adds Enterprise AI Credit Controls and Admin APIs (UX News, October 2026), and Manage AI credit limits with SCIM (Figma, October 2026).

Victorino Group helps finance and engineering leaders design AI spend controls tied to people, approvals, and identity. Let’s talk.