GPT-4's 8K API charged $60 for a million output tokens in 2023. A model that matches it on most benchmarks now runs for a fraction of that. Enterprise AI spending rose anyway.
Agentic workloads explain the gap. A coding or research agent burns roughly 1,000 times the tokens of a chat reply, and the same task can swing 30-fold between runs. Falling unit prices met rising volume, and volume won.
Gartner's November 1 note draws the operational conclusion: comparing models on per-token cost "ignores the key cost drivers of token consumption and success rates." Benchmark cost per successful task instead.
| Signal | Number | Source |
|---|---|---|
| Agentic vs chat token use | ~1,000× | arXiv 2604.22750 |
| Cost per task, orchestration swap | −41% ($0.21 → $0.12) | arXiv 2607.06906 |
| Token share of variable run cost | 20–25% | McKinsey |
Three numbers that reframe AI spend, 2026
Why the meter stopped matching the bill
Token pricing was designed for single-turn inference. One prompt, one answer, one bill.
An agent works differently. It re-reads its own context at every step, calls tools, drafts, checks, and retries. Input tokens carry the cost, and the context window becomes a recurring expense rather than a one-time input.
The variance is the part finance teams miss. Bai and co-authors at Microsoft Research and Stanford ran eight frontier models on SWE-bench Verified. Runs on the same task differed by up to 30 times in total tokens. Accuracy peaked at intermediate cost and flattened above it.
More tokens did not buy more correct answers.
Frontier models also cannot forecast their own usage. The best correlation reached 0.39, and every model underestimated real cost. A budget built on average token price will drift the moment agent behavior changes.
Cost behaves as a distribution. Budgets built on a point estimate do not survive it.
The harness moves the bill more than the model
A July study put the orchestration layer under a controlled test. Twenty-two locked evaluation tasks, six foundation models, one variable changed: a frozen production loop against a purpose-built agent harness.
The harness cut blended cost per task 41%, from $0.21 to $0.12. Median wall-clock time fell 44%, from 48 to 27 seconds. Tokens per task dropped 38%, from 14,200 to 8,800. Task-completion quality held at parity, 0.78 to 0.81.
Every model got cheaper, by 33% to 61%. Quality gains tracked baseline model strength almost perfectly, with a correlation of 0.99. Quality per dollar rose 82%. Task completions per million tokens climbed from 54.9 to 92.0.
The authors state the finding plainly: on this workload, the orchestration layer moved cost per task more than the full spread of the model menu did.
That inverts the usual procurement conversation. If the harness dominates, then model differentiation compresses and the buyer's bargaining power shifts to caching, routing, context discipline, and failure governance.
The model is becoming a component. The harness is becoming the product.
The harness result also explains why model launches feel less decisive than they used to. A stronger model raises the ceiling on hard tasks, but the same orchestration discipline applies to every model an organization runs. The multiplier travels with the harness, not with the checkpoint.
Where the hidden spend accumulates
Refinement dominates the bill. McKinsey cites an analysis of production agentic coding workflows that puts roughly 60% of an agentic task's cost into reviewing and refining answers. Creation is the cheap part. Verification is the expense.
Input tokens drive that pattern. The Microsoft and Stanford study found input, not output, carries overall cost, because every step replays the context built so far. A long session pays for its own history at each turn.
Prompt caching lowers the effective input price, but only while the context shape stays stable. Change the prompt prefix and the cache misses, so the discount disappears at the moment the agent starts improvising.
Cache hit rate belongs in the same conversation as model choice. Two teams can run the same model at the same list price and see effective input costs that differ by multiples, purely on how they shape context.
Failure spend compounds the rest. A task that fails and reruns costs twice. A loop that retries five times costs six. Governance that caps iterations and stops runaway loops cuts average spend per task sharply in vendor benchmarks; treat those vendor numbers with care, but the mechanism is not in dispute.
McKinsey's banking example shows how the pieces stack. A completed account onboarding needs five to seven agents, several deterministic systems, and two to four teams providing oversight. Each handoff replays context. Cost compounds with workflow complexity, not with the number of users.
This is why per-seat pricing broke first in customer service. Seats scale with people. Agent cost scales with the shape of the workflow.
What a resolved task actually costs
McKinsey's agentic economics work puts the token line in perspective. For a customer-service agent in banking, token costs are 20% to 25% of variable run costs. Human oversight accounts for 70% to 75%.
The absolute numbers are large enough to matter at portfolio level. A single-agent banking workflow can cost $20,000 to $30,000 a year to run. A multi-agent team can reach $100,000 to $200,000. A conversational onboarding agent handling 2,500 customers a year costs $10,000 to $15,000.
Fully loaded, the same onboarding workflow can fall from $50–$150 per customer to $10–$30. The saving is real. It comes from workflow redesign, exception handling, and fewer human reviews, not from a cheaper token.
As we wrote in October, the data-center refresh trap is real: cheaper tokens and bigger bills arrived together. What changed since is the measurement standard.
The market is repricing in public
Customer-service vendors moved first because their unit of work was already countable.
| Pricing unit | Example | What it meters | Exposure |
|---|---|---|---|
| Per seat | Legacy helpdesk | Headcount | ✗ Shrinks as agents resolve |
| Per token | Model APIs | Compute consumed | ✗ Rewards verbosity |
| Per resolution | Intercom $0.99, HubSpot $0.50, Zendesk $1.50 committed | Completed work | ◐ Definition of "solved" |
| Per verified outcome | Zendesk tiers, Salesforce Agentforce | Quality-gated result | ✔ Aligns vendor and buyer |
List prices as reported by TechTarget and vendor pages, 2026
The weak point sits inside the word "solved." Zendesk counts 72 hours of silence as a resolution. Other vendors use no-reply, no-escalation, or no-repeat windows as stand-ins. There is no shared standard, so the prices are not comparable.
The measurement problem is now the pricing problem.
Salesforce paid roughly $3.6 billion for Fin, formerly Intercom, after a June agreement and a September 10 close. Fin reports an average end-to-end resolution rate of 76% across more than 30,000 customers. Salesforce then agreed to buy Cresta on November 2, its tenth acquisition agreement of 2026.
The future of customer service is autonomous, intelligent, and built on trust.— Marc Benioff, chair and CEO, Salesforce
The strategic logic is straightforward. Per-seat licensing is exposed to agent-driven headcount declines. Owning the agent captures the outcome fee instead.
How to underwrite an agent company
Five numbers separate an agent business from an agent demo.
Resolution rate on a defined task set, not a demo set.
Cost per resolved task, including failed and retried runs.
Tokens per task trend across harness releases.
Human oversight share of variable cost.
Pricing unit, and the audited definition behind it.
Two of those numbers are diagnostic. A falling cost per resolved task with a flat resolution rate means the harness is improving. A falling cost per token with a rising cost per task means the agent is doing more work to reach the same place.
The second pattern is common. It also explains why token-metered inference resellers are short the efficiency curve. Every harness improvement reduces the tokens a customer needs, and their revenue is the token.
Orchestration, observability, and outcome-priced agents sit on the other side of that trade. They earn more when the same task costs less to finish.
A buyer can manage the downside. Demand the resolution definition in writing, sample the failures, and price the fee against a measured baseline rather than a vendor case study. The discipline that made cloud unit economics work applies here.
The repricing also changes what a multiple is measuring. A seat-based software company is valued on seats times price. An outcome-priced agent company is valued on resolved volume times fee, minus the cost of the harness, the oversight, and the failures. Different margin curves, same support queue.
The counterargument deserves a hearing. Outcome pricing can hide quality. Vendors choose the stand-in for "resolved," and the stand-in can flatter the vendor. Buyers who accept a resolution fee without an audited definition are buying a metric, not a result.
What to watch next
Gartner's benchmark framing will move into procurement language first, because software engineering leaders already control the model-harness budget. Expect cost-per-successful-task to appear in vendor scorecards before it appears in audited financials.
The resolution fee itself is converging. Intercom at $0.99, HubSpot at $0.50, and Zendesk at $1.50 committed bracket a narrow range for a solved contact. A market that settles on a price will eventually demand a definition.
Cost-per-successful-task enters standard procurement templates.
An audited definition of "resolved" from a major vendor or standards body.
Harness-level token efficiency reported beside model benchmarks.
Outcome-priced revenue disclosed separately from seat revenue.
Orchestration, not model choice, is the largest controllable cost lever.
Revenue metered on tokens is short the efficiency curve; outcome-priced agents are long it.