Outcomes Pay In Tokens

On May 14, 2026, Anthropic told its subscription users that beginning June 15, agent traffic would no longer draw from the same usage pool as a chat session. The Agent SDK, the claude -p subprocess path, the GitHub Actions runner, and every third-party tool that authenticates through the Agent SDK would get their own credit envelope. Twenty dollars for Pro, one hundred for Max 5x, two hundred for Max 20x. Billed at standard API rates after that. Weeks later, the company quietly reverted the page. Pause noted. Status quo restored.

The reason for the announcement was a sentence Anthropic had stopped denying. Power users were running agents through the subscription path and getting fifteen to thirty times the value of the API ticker on the same compute. The fifteen-to-thirty multiplier was a billing seam Anthropic could not afford forever in a price war with OpenAI, which on the same day, May 14, had handed Codex Pro to new business customers for two months free.

A subsidy of fifteen to thirty times is a structural admission about the economics of AI agents. No marketing department writes that number on purpose. Whatever the buyer pays at the top of the funnel, the actual ledger of an agentic workload is denominated in tokens at the bottom. And tokens cost cash.

This is the thing executives who buy AI agents the way they used to buy SaaS will get wrong this year. They will sign a per-seat or per-outcome contract that looks clean on paper, and they will discover that the company behind it is running on margins that no SaaS finance team would recognize. Or worse, they will become that company.

Twenty-Three Cents Of Every Dollar

ICONIQ Capital published its bi-annual State of AI in 2026. The data inside it has been the most-cited block of numbers in any board deck I have read this month. Average gross margin for AI-native products in 2026 is fifty-two percent, up from forty-one in 2024 and forty-five in 2025. Mature SaaS sits at seventy-five to eighty-five. The gap is twenty-three to thirty-three points.

Labor is the line that fell. Talent costs at scaling-stage AI companies dropped from thirty-two percent of spend to twenty-six. The line that rose is inference. Inference now eats twenty-three percent of revenue at the median scaling-stage AI company. Eighty-four percent of those companies report six or more points of gross-margin erosion attributable to model and infrastructure cost alone.

Set that down next to the rest of your portfolio. If you own a SaaS line where gross margin matters, you already know that ten points of margin compression is the difference between a growth story and a cash story. Twenty-three points of revenue, paid directly to a model provider, is a cost line you have to architect around. Optimizing on the margin will not close it.

Semiconductor providers are delivering inference cost reductions of sixty to seventy percent per year. That sounds like a tailwind. Run the second derivative. Per-token prices are falling. Per-workload token consumption is rising at roughly the same pace. The aggregate spend line does not bend. The companies that take the unit-cost decline as gross margin are the ones writing their own inference stack and routing rules. The companies that take it as more loops at the same dollar are the ones whose CFOs are reading ICONIQ's report and wondering why their AI line gets bigger every quarter.

Outcome Pricing Hides A Distribution

Intercom charges ninety-nine cents per Fin AI resolution. Zendesk charges a dollar fifty on committed volume and two dollars on pay-as-you-go. Sierra publishes nothing, but third parties peg an annual floor at one hundred and fifty thousand dollars with setup fees of fifty to two hundred thousand stacked on top. The shape of the contract is the same across all of them. The buyer pays when the agent resolves a customer issue. No resolution, no charge.

This is being sold as the cleanest deal in software. For the buyer, it usually is. Outcome-based contracts compress procurement risk into a number a finance team can model. The vendor only gets paid when the workflow ends in the customer's win. Sierra built its entire business model on this.

Look at the seller's side of that ledger. The price per resolution is fixed. The cost per resolution varies. Easy cases finish in two model calls and a database write. Hard cases loop. They read knowledge bases, they fetch order histories, they escalate, they retry, they call tools, they reason through edge cases. The token bill on a hard case can be ten times the token bill on an easy one. Sometimes more.

A ninety-nine cent resolution is the expected value of a cost distribution, dressed as a single number. If the vendor's loop is bounded and routes hard cases to a small model first, the distribution is tight and the gross margin lives. If the loop is unbounded or escalates lazily to the most expensive model the vendor has on tap, the long tail eats the contract. The vendor sells one resolution at a profit, two at break-even, and the third at a loss. The buyer never sees it because the bill is flat. The vendor's income statement absorbs the variance.

This is the unit economics structure of an insurance company. SaaS finance teams do not have the muscles for it. The premium is paid up front. The claim is whatever the loop costs. Solvency depends on the law of large numbers and on knowing what the actual cost distribution looks like before you set the price. Most AI agent vendors do not have a year of historical loop data on their own production traffic. They priced before they knew.

Inference Efficiency Is The New Gross Margin

If you are running an AI product line, the lever that will decide whether you have a business in eighteen months is not the brand of the model you ship on. It is the routing.

ICONIQ found the top-margin AI companies share one operational fact. They route the majority of incoming tasks to small, cheap models and only escalate the complex residual to a frontier model. This is the inference efficiency ratio. It is the AI-era equivalent of COGS over revenue, and it is becoming the single most-discussed metric in AI CFO circles.

A small builder sees this as a model selection problem. A real one sees it as a control surface. The router is a workflow that decides, for every incoming task, which provider to call, which context to attach, which cache to hit, whether to use a structured tool path or an open reasoning path, and when to stop. The companies with the cleanest router are the ones whose inference line did not double when their volume did. The companies without one are the ones writing apology emails to investors about cost of goods sold.

There is a second-order effect that does not get discussed enough. The router is also the place where the model swap happens. Every six months a smaller, faster model becomes good enough to take a task that used to belong to the largest one. If your router is real, that swap is a config change and a margin lift. If your router is a switch statement that someone wrote during the pilot and never touched, that swap is a six-week project that gets prioritized after the next feature ships, which is never.

A two-order-of-magnitude per-task cost gap between DeepSeek V4 Flash at fourteen cents per million tokens and a frontier reasoning model at fifteen dollars per million is not a rounding error. It is a hundred-to-one ratio. Routing is where that ratio gets monetized. Or wasted.

What The June 15 Almost-Change Told Buyers

Anthropic pulled back the Agent SDK billing split. The reason they had to announce it in the first place is the part the buyer should keep on the desk.

Pricing in this market is unstable because cost-to-serve in this market is unstable. The provider charging you a flat subscription for agent traffic today is doing it because a competitor's subscription is doing the same thing and pulling power users across. The provider charging you per outcome today is doing it because the buyer wants the alignment narrative. Both of those framings can survive a competitive cycle. Neither of them survives a year of unit-economic reality if the underlying loop cost is out of control.

If you are a buyer, you have three weeks to a quarter of pricing stability with any agent vendor before the contract structure changes underneath you. The vendor will frame the change as an upgrade. It is a margin repair. The careful buyer reads every agent contract with the assumption that the per-outcome price will rise, the credit envelope will shrink, or the definition of "resolution" will tighten in the next twelve months. Cap the per-incident exposure. Demand transparency on what counts as a resolved outcome, and ask whether the vendor's router lives in code you can audit or in a black box you cannot.

If you are a builder, the implication is the opposite one. The companies who ship now without an inference cost model will sign contracts they cannot honor at scale. The ones with a real router will be repricing upward into a market that has not yet learned how to read the bill.

The Strategy That Wins This Decade Owns The Loop

Here is the practical truth I keep landing on with the executives I work with at Agor AI Advisory. The economics of AI agents in 2026 do not reward owning a model. They reward owning the loop the model runs inside.

The loop is where the prompt budget lives. The loop is where the cache decisions get made. The loop is where the small-model-first routing logic sits. The loop is where the timeout that prevents an unbounded reasoning chain gets enforced. The loop is where the eval harness measures, on real production traffic, whether a switch to a cheaper model just cost you accuracy or saved you margin.

A company that owns its loop has a router it can change in a day, a cost ceiling it can defend in a quarter, and a margin trajectory that compounds as inference prices fall. A company that buys an off-the-shelf agent and lets a vendor own that loop has rented its cost structure from someone whose margin is the inverse of its own.

This is the asymmetry that off-the-shelf vendors do not advertise. When the model swap happens, the vendor captures the cost reduction. The buyer keeps paying the same flat per-outcome price until renewal. Sometimes after.

What Buyers Should Actually Ask

I am putting two questions on every agent procurement checklist this quarter.

The first one is mechanical. What is the vendor's average token cost per successful outcome over the trailing ninety days of production traffic, and what is the ninety-fifth percentile? If the vendor cannot answer that, the vendor does not have a router. The vendor has a wrapper.

The second one is structural. If the underlying model price drops by half tomorrow, where does the dollar go? The buyer's contract, the vendor's gross margin, or some negotiated mix. Get that in writing now, while the seller still wants the logo. The default for every contract I have read this year is that the seller keeps the windfall. That is the default. It does not have to be the result.

A Quiet Test For The Pilot

There is a third test I run during pilots. I ask the vendor to send me one week of cost-per-resolution data, bucketed by intent. If the answer comes back as a single average, the vendor is operating on hope. If the answer comes back with a histogram and an outlier list, the vendor has built the company on the right axis.

Most of the well-funded outcome-priced agent companies in the market today fall into the first bucket. They have raised money on the alignment story without building the operational machinery the story requires. The buyer who knows to ask the histogram question gets to see, in five minutes, which side of the next twelve months that vendor is on.

Build It Where The Margin Lives

A consulting tool that you bought from a vendor will optimize for the vendor's margin. The router will favor the model that pays the vendor the highest revenue share. The eval suite will measure quality and stop short of cost. The cost ceiling will be a slider in the vendor's admin panel that nobody on your side has access to.

The version of this story that ends with your company on the winning side of the gross margin gap is the one where the loop is architected and owned inside your walls. The model can be anybody's. The router has to be yours. The cost data has to be yours. The eval harness that compares incoming model releases on your real production traffic has to be yours.

This is the work Agor AI Advisory does. We architect the agent loop, the router, the cost ceiling, and the eval harness as a single integrated system, sized for your business, owned by your team, instrumented from day one with the cost metrics that the vendor pricing pages do not show you. The buyers who win the next eighteen months are the ones who treat the inference line the way they treat the payroll line. A real line item, with a real owner, with a real plan to bring it down each quarter.

The token tax is real. The router decides who pays it. Architect the loop, or rent it from someone whose margin depends on you not noticing the bill.