The economics of agentic AI are diverging: the cost of a single token is collapsing while the cost of running an agent barely moves. Three moves this week are one restructuring seen from three angles. Anthropic launched Haiku 5.5 at $0.10 per million input tokens and $0.50 per million output on Oct. 7, collapsing the small-model floor. Mistral’s ML4 at $0.68 per million input and $2.09 per million output landed a day earlier, undercutting the proprietary one. And Google moved its free tier down to Flash-Lite while documenting compute-based usage limits that factor in prompt complexity, the models and features used, and chat length – refreshed every five hours against a weekly cap.
What is moving is the unit of account. As raw inference commoditizes, the bill is shifting from token counts to compute, effort and – rarely, so far – outcomes. Google now meters compute directly: the buyer pays for the intensity of the work, not the volume of text. That is the direction the market is walking, one layer at a time.
The hidden cost sits in the refinement loop. McKinsey’s July 2026 report, drawing on 20 production agentic coding workflows, puts about 60 percent of an agentic task’s cost in response refinement – the cycles of checking, correcting and re-verifying outputs – rather than the initial inference call; the underlying study (Salim et al., arXiv 2601.14470) measured a 59.4 percent review-and-refinement share. Ratios vary by workload and verification depth, and this is consultancy analysis over a small coding sample, not a universal constant. But it explains why cheaper tokens have not produced cheaper agents. In McKinsey’s Enterprise AI FinOps Survey (May 2026, 75 respondents – a small sample), 93 percent of organizations exceeded their AI budgets, with refinement cost cited as a key reason the forecast missed the actual.
Vendors are repricing around that volatility. Orb’s 2026 pricing study of 80 AI agent companies – vendor research; Orb sells billing infrastructure – finds 95 percent on hybrid pricing, up from 92.4 percent a year earlier, 91.3 percent metering usage, and pure outcome pricing rare at 3.8 percent and falling. The unit climbing fastest is effort: compute time, number of steps, task complexity. Stripe, cited in the same study, finds 92 percent of usage-charging AI businesses have already changed their pricing at least once. Repricing is the expected case, not the exception.
The line between record and projection matters here. BCG’s research frames dynamic token pricing – inference prices moving with supply and demand like wholesale electricity – as the likely next evolution, with up to $140 billion in annual “dark value” stranded in compute supply-demand mismatches. That is consultancy projection, not current practice. Deloitte, citing Gartner, projects at least 40 percent of enterprise SaaS spend could move to usage-, agent- or outcome-based models by 2030; Gartner separately projects generative-AI cost per resolution in customer service passing $3 by 2030, above many offshore human agents. Vendor list prices sit below that today – Intercom at $0.99 per resolution, Gorgias at $0.90 per resolved interaction – which means vendors publishing those rates are pricing against a cost base that may not stay where it is.
For buyers, the practical shift is in what gets watched. Audit the unit on the invoice – tokens, compute, steps, resolution – because metering opacity is part of the price. Expect the rate card to move: 92 percent of usage-charging vendors have already repriced. Track cost per task or per resolution rather than per-token list price, and treat any vendor’s definition of a billable outcome as a contract term, not a detail. Even the week’s other headline points the same way: OpenAI’s revenue reading roughly $20 billion lighter than reported, a gap our analysis traced to accounting methodology rather than demand. The price of intelligence is falling. The price of the arrangement around it has not been written yet.