Cheaper Tokens, Pricier Outcomes: The New Unit Economics of AI

On 22 September, Anthropic cut the list price of its top Opus model by a fifth, to $4 per million input tokens and $20 per million output tokens1. Eight days later, Google launched Gemini 4 Argon at an introductory price of $2 and $10, half of the $4 and $20 it says will apply once the introductory period ends3. In the same weeks, Gartner put its forecast for worldwide AI spending in 2026 at $2.7 trillion, up 49.5% on the year45.

The unit got cheaper. The bill got bigger. For anyone who signs off on AI budgets, the gap between those two facts is where the money goes.

The price of a token is not the price of the work

Per-token prices are real and falling. Anthropic says its new model costs 40% less than its predecessor on typical workloads at default settings1. Its price list also halves the rate for asynchronous batch jobs and charges cached input at a small fraction of the base price2. Google prices cached input for Argon at 95% off3.

Yet about one in five respondents to McKinsey's 2026 State of AI survey say AI operating costs, token costs included, have constrained their use of AI67.

Our read: the reason is volume. A chat answer uses a few thousand tokens. An agent that plans, calls tools, reads documents, checks its own work and retries can use many times that for a single task, and it runs whether or not a person is watching. Cheaper units multiplied by far more units can still produce a larger invoice. Price per token is an input cost, like the price of steel. Nobody runs a plant on the price of steel alone.

Profit is the stubborn number

McKinsey's survey, fielded from 4 May to 8 June 2026 with 1,719 respondents, found that 37% say AI has contributed to their organization's EBIT, essentially unchanged from a year earlier678. About 6% qualify as AI high performers, also flat78. Meanwhile the share of companies with more than $1 billion in revenue that are scaling AI agents rose from 27% to 40%78.

More agents, the same share seeing profit. In June 2025 Gartner predicted that more than 40% of agentic AI projects would be canceled by the end of 2027, citing escalating costs and unclear business value910.

From seats to units of work

Software pricing is shifting under buyers' feet. Gartner estimates that up to $234 billion of enterprise application software spend, about a fifth of enterprise SaaS spend, is at risk from agentic AI by 20301112. Its reasoning is that seat-based pricing assumes more users means more revenue, and agents break that link when they do work people used to do in application screens1112. Gartner advises vendors to move toward value based on outcomes rather than interfaces11.

Vendors are already moving. SAP, for example, prices Joule Agents on AI Units consumed per action, with no separate charge for the Joule assistant13.

Fig. 01Art · Economics

That changes how to read a price sheet. Five questions:

• What is the unit? A token, an action, a credit, a conversation and a "resolution" are not interchangeable. Ask for a worked example on one of your real tasks.
• What triggers a charge? Retries, tool calls, failed runs and background monitoring may all consume units.
• What is the price after the promotion? Argon's introductory rate doubles when the period ends3. Model your business case on the standard rate.
• Which discounts can you actually use? Batch and caching discounts are large23, but only for work that can wait or that reuses the same context.
• Who controls the ceiling? Insist on spend caps, alerts and per-workflow reporting, so a looping agent cannot spend a quarter's budget over a weekend.

Measure the task, not the token

Measure where profit is made: the finished task. Three numbers do most of the work.

Cost per resolved task. Add up model and platform charges, the cost of human review time, and the cost of fixing errors. Divide by the number of tasks the system completed without being handed back to a person. Compare it with what the same task costs today.

Rework rate. The share of AI outputs a person had to correct or redo. A low token bill with a high rework rate is a hidden labor cost.

Cycle time. Elapsed time from request to done, including time waiting for review. If approvals still queue for two days, a model that answers in two seconds has not changed the outcome.

When to stop a pilot. Write the stop rules before you start. Our suggestions: stop or redesign if, after an agreed run of real volume, cost per resolved task is not trending below the baseline; if rework is not falling; if the business case only works at an introductory price; or if no one in the business has agreed to own the process once the pilot ends. A pilot that cannot fail is not a test.

The forward-deployed question

On 29 September, Gartner predicted that by 2028, 70% of enterprises will abandon agentic AI built through vendor forward-deployed engineering (FDE), the model in which a provider's engineers work directly with a customer to build and deploy its agents141516. As reported, Gartner's reasoning is that costs rise and customers cannot evolve the systems without outside help1516. It also predicted that through 2028, fewer than 20% of FDE engagements will turn recurring customer needs into capabilities in the vendor's core product, and warned of "FDE washing," where ordinary consulting is relabeled and sold at premium rates1516. Gartner analyst Mukul Saha is reported as saying that success depends on engagement structure: scope, incentives, governance, ownership and exit1516.

Olympas's Management Consulting practice includes Forward Deployed Engineers, so we have a stake in this, and readers should weigh our view accordingly.

What Gartner gets right: dependency is the failure mode. A system that only its builders can change is a liability, however well it works on launch day. Relabeled services at premium prices are a real risk too.

Where our view differs: our read is that the problem is not having engineers on the client's floor. It is engagements designed never to end, or designed to feed a vendor's product roadmap rather than the client's process. Proximity to the work is often what makes an agent fit real exceptions. The question is who owns the result.

What a buyer should demand from any forward-deployed engagement, including ours:

• Measurable outcomes, agreed up front. Cost per resolved task, rework rate and cycle time against a measured baseline, with fees tied to milestones.
• Ownership of the system. Code, prompts, evaluation sets, configurations and data pipelines belong to you, in your repositories and your cloud accounts.
• Your people in the work. Named internal staff who pair with the outside team from the first week, not at the end.
• A dated handover. Runbooks, documentation and a period in which your team runs the system while the outside team only advises.
• An exit test. The engagement is done when your team can change, retrain and retire the system without calling anyone.

What to do now

• Pick your three highest-volume AI use cases and calculate cost per resolved task, rework rate and cycle time for each, starting with a baseline measured without AI.
• Ask each AI vendor for a worked cost example on one of your real tasks, at the standard price, with retries and tool calls included.
• Set spend caps and alerts per workflow, and route routine tasks to cheaper models or batch processing where the work can wait.
• Write stop rules for every running pilot this month, including a named business owner for the process after the pilot.
• Before renewing any seat-based software, ask the vendor how pricing will work when agents, not people, do the work.
• Review every outside AI engagement for ownership, knowledge transfer and an exit date. If any of the three is missing, add it before the next invoice.