Reality Check / Reporting
Count finished work, not tokens: a sponsored pitch, read closely
Judge an AI tool by cost per finished, reviewed result, not by usage. That is the argument in a sponsored HPE piece. Its 30 to 50% savings figure is the vendor's own.
AI-generated editorial illustration · EachLabs
A good question can still arrive inside an ad. That is the case with "Stop counting tokens, start counting what ships," which MIT Technology Review published on 2 November 2026. The page is labeled sponsored and provided by HPE. It states that it was produced by HPE and not written by the magazine's editorial staff. So read it as a vendor's argument, not as reporting.
What the piece says
The core point is about how AI coding agents work. An agent does not answer once. It plans, writes code, runs tests and repeats to fix errors. The piece says a single task can use 10 to 50 times more tokens than a basic chat prompt. With per-token pricing, it argues, the most valuable workflows can become the most expensive.
It cites pressure that is already visible. It says Uber reportedly used up its entire 2026 AI budget by April after expanding use of a coding agent. It says GitHub Copilot moved from flat-rate pricing to metered credits. It also says Gartner predicts AI coding token costs will pass average developer salaries by 2028. These are the sponsored piece's statements. We have not checked them against the original sources.
The useful idea
The piece says caps and approvals "optimize the wrong number." A token count measures consumption, not results. Some spending is waste, such as retries and runaway loops. Some is the cost of getting work done. A meter cannot tell them apart.
Its alternative is cost per merged pull request. That is total AI-assisted development cost divided by the number of changes that pass review and are merged. It says to track rework and defects alongside, so cheaper changes do not hide lower quality. It also suggests tracking cost per active developer, because averages across all licenses hide heavy users.
On usage, the piece says the top 5% of users generated most of the consumption. It says the heaviest user consumed roughly 450 times more than the lightest on the same license. It does not say whose data this is. Treat it as an illustration, not a benchmark.
Where the pitch begins
The piece then recommends routing routine work to open-weight models a company runs itself, and keeping frontier models for hard problems. It says HPE did this for more than 150 internal applications and cut token costs by 30 to 50%. That figure is HPE's own, reported in HPE-produced content. The piece gives no method and no independent check. It sits next to a plug for HPE Private Cloud AI.
To its credit, the piece hedges. It says benchmarks show leading open-weight models can approach frontier performance on some tasks, but that does not prove equivalence in every codebase. It says owned capacity has real costs: power, operations, maintenance and depreciation. It says low or unpredictable usage favors paying per call. It advises measuring both metrics before changing infrastructure, and staying put if they do not improve.
Those caveats are fair to include. They also leave the savings figure as a vendor claim until someone outside HPE shows how it was measured.
The takeaway
You do not need a data center to borrow the idea. Before judging any AI tool by its bill or its usage, pick the unit of finished work and divide by that. This list is our reading of the piece's logic, not advice the piece gives for small teams.
- Define "done." For code, the piece says it is a merged change. For a small business, it might be a published post or a resolved support ticket.
- Count only work that passed human review. Drafts that get thrown away are cost, not output.
- Track rework next to cost. A cheap result that needs redoing is not cheap.
- Baseline first. Measure before you switch tools or vendors, then compare.
- Discount vendor savings figures. A number from the seller's own deployment is a lead to test, not a result to bank.
The pitch says to count what ships. Fair enough. Count it for the product being sold, too.
Source
Original report · Published 2026-11-02 13:10:32 UTC. Published under this site’s editorial settings.
This article was drafted with AI and published under the site’s editorial settings.
