The FinOps Foundation published a token economics framework in June 2026 that recommends four unit metrics for making AI spend legible to finance. Not one of them is cost per token. Cost per query, cost per user per month, cost per workflow completion, cost per business transaction: those are the numbers the framework says belong on a shared dashboard. The token price, the figure every provider leads with and every FinOps deck still tracks, does not make the list.
That gap is the whole story of AI cost management right now. Teams instrumented themselves to watch the one number that has become the least useful predictor of what they will actually pay.
Why the token became the default number
Tokens won by default because they were the only unit the bill exposed. When an OpenAI or Anthropic invoice arrives, it is denominated in tokens. So the first generation of AI cost dashboards did the obvious thing: they counted tokens, multiplied by the rate, and called it observability. For a chat feature where a user types a question and gets an answer, that was close enough. One interaction, a few thousand tokens, a predictable cost.
Agents broke the arithmetic. An agentic workflow does not make one model call. It plans, retrieves documents, calls tools, validates its own output, and retries when the validation fails. The Forbes analyst Janakiram MSV laid out the real cost equation in July: total spend equals task volume, times attempts per task, times tokens per attempt, times effective token price, plus tool and infrastructure costs. The token price is one of five terms, and it is the term that has been falling while the other four climb.
His illustration is worth keeping in your head. A twenty-fold increase in tokens per task combined with a seventy-five percent price cut still produces a five-times-higher model charge. Goldman Sachs projects token consumption multiplying twenty-four times between 2026 and 2030. Prices per token can keep dropping and your AI bill can quadruple in the same year, and both statements will be true. That is the same trap I wrote about in why token prices drop but the bill keeps climbing, now measured against agents instead of chat.
Meanwhile the price signal itself got noisier, not clearer. In one eight-day stretch this July, three major models launched: Grok 4.5, GPT-5.6, and Meta’s Model API. Meta priced at $1.25 per million input tokens and $4.25 per million output. OpenAI’s GPT-5.6 Sol landed at $5 input and $30 output. If your dashboard tracks a blended cost-per-token figure across a fleet that just re-shuffled which model handles which task, that figure now describes your routing decisions more than your efficiency.
The four numbers that actually mean something
The metric that replaces cost per token is cost per outcome, and the reason is not philosophical. It is that an outcome is the smallest unit your CFO can attach a value to. A resolved support ticket has a dollar value. A summarized page has a value. An accepted code suggestion has a value. A token does not.
The practitioner playbooks that emerged this year converge on a small set of these. The usage.ai KPI playbook frames cost per inference (total inference cost divided by number of requests) as the base unit economic metric for production AI, then layers utilization and coverage on top: GPU utilization above fifty percent for inference, commitment coverage between sixty and eighty percent for stable workloads. Vantage frames the same idea as cost per output, and its examples are the useful part: for a customer support AI, cost per resolved ticket; for a coding assistant, cost per accepted suggestion; for a document tool, cost per summarized page.
The through-line is that you divide fully loaded cost by verified results, not by activity. Activity metrics reward the wrong behavior. A model that retries ten times and eventually fails logs more tokens than one that succeeds on the first try, and a token dashboard will show the failing model as the bigger line item without ever telling you it failed. Cost per verified outcome, fully loaded cost divided by outcomes that actually happened, is the version that survives contact with an agent.
None of this is exotic to anyone who has run an operations business on unit economics. When I ran costs in Saskatchewan telecom years ago, nobody on the finance side cared about cost per minute of switch time. They cared about cost per subscriber and cost to acquire, because those tied to revenue. The engineering teams tracked the infrastructure units and the finance teams tracked the outcome units, and the entire job of the person in the middle was translating one into the other. AI FinOps is that same translation job with tokens standing in for switch minutes. The teams that struggle are the ones still publishing the switch minutes to the CFO.
The harness is 40 to 60 percent of the bill
Here is the part most token dashboards miss entirely. The FinOps Foundation’s token economics working group, in its June 2026 material, states plainly that the infrastructure surrounding the model call routinely represents forty to sixty percent of total feature spend. Vector databases, caching layers, reranking, code execution sandboxes, observability, human review. The framework’s warning is direct: cost-per-query figures that exclude the harness will systematically underreport.
So a cost per outcome that only counts tokens is not just incomplete, it is misleading in a predictable direction. It always makes the feature look cheaper than it is, which means every ROI case built on token-only numbers is quietly overstated. If you are going to move to outcome metrics, you have to pull the surrounding infrastructure into the numerator, and that infrastructure lives in your regular cloud bill, tagged (or not) alongside everything else. This is one reason the AI gateway pattern matters beyond routing: it is one of the few places you can capture request-level telemetry and start associating a model call with the rest of the stack it triggered.
The optimization hierarchy that falls out of this is worth stating. The single highest-leverage move for most organizations is not negotiating a lower token rate. It is model right-sizing, matching the task to the cheapest model that clears the quality bar, which the FinOps Foundation puts at sixty to ninety percent potential savings. You cannot make that decision from a token dashboard. You can only make it from a cost-per-outcome view that shows which model produces an acceptable result for the least fully loaded cost.
Attribution is the hard part, and providers do not help
The reason more teams have not already made this shift is that the raw data fights you. Model providers do not natively support the tagging structures FinOps teams rely on. The FinOps Foundation’s own guidance sequences the workaround: API key governance first, then provider-native features, then a proxy layer for organizations that need cross-provider attribution.
Provider granularity varies more than most buyers realize. Vantage’s assessment is that Anthropic and Cursor break costs down to the API key or developer level, which is immediately usable for attribution and anomaly detection. OpenAI requires additional work. AWS Bedrock loses developer-level visibility entirely when it is purchased through the marketplace. So two teams running identical workloads on different providers can have wildly different abilities to answer “what did this outcome cost,” through no fault of their own instrumentation.
The deeper problem is that agents cross business contexts inside a single run. One agent execution might touch a support workflow, a billing lookup, and a document generation step. Traditional tagging assumes one resource maps to one owner, and that assumption breaks the moment a single agent run serves three business processes. The practical answer that has emerged is call-level metadata: every model call carries the feature, team, and business process it serves, stamped at the application layer, not inferred from the bill afterward. That is application work, not a FinOps dashboard setting, which is exactly why it keeps getting deferred. It is also the difference between a cost-per-outcome number you can defend and one you made up. The certification ladder that FinOps built around this exists precisely because the skill is now half engineering.
Where to start this quarter
If your AI reporting today is a token counter, the move is not to rip it out. Tokens still matter as an efficiency input; they just make a terrible headline number. Pick one AI feature that finance keeps asking about. Define its outcome in business terms, a resolved ticket, a shipped pull request, a processed invoice. Then build the numerator honestly: model tokens plus the harness around them, the vector store, the retries, the review time. Divide by outcomes that actually completed, not attempts. Publish that one number next to the token chart for a month and watch which one the CFO reads.
The FinOps Foundation’s maturity model calls this crawl, walk, run, moving from account-level visibility to workload attribution to request-level unit economics. Most teams are stuck at walk because run requires the application-layer metadata nobody wants to prioritize. But the teams that get there stop having the argument about whether AI is worth it, because they can answer the only question that ever mattered: what does one good outcome cost, and is it less than the outcome is worth. That is the same question a good AI forecasting practice is trying to answer forward in time, and you cannot forecast an outcome cost you have never measured.
Cost per token was the number the invoice handed you. Cost per outcome is the number the business was always asking for. The gap between those two is where most of 2026’s AI budget overruns are hiding.