Twenty pence looks cheaper than forty-five until one job needs several tries. Imagine two AI systems preparing the same document. The less expensive attempt can become the more expensive accepted result, and that comparison still has human review ahead of it. The useful number sits at the end of the work.
OpenAI’s September 8 essay The Work Now Within Reach connects model capability, distribution and computing infrastructure to work becoming economically practical. Its author, Sarah Friar , was writing as OpenAI’s chief financial officer. The completed task is the part of that corporate argument a customer can turn into a concrete comparison.
A model request can finish while the job remains unfinished. Someone may still need to check a citation, reconcile a claim with its source or repair the file before use. Count the accepted result and those steps become visible. Count only requests and some of the most demanding work can disappear from the comparison.
Define acceptance first, then follow every attempt needed to reach it. That turns a broad cost proposition into something a team can inspect for its own work. The same system may be useful on one task and expensive on another, depending on the failures and repairs each produces.
Six accepted files out of ten changes the arithmetic
Suppose a team asks an AI system to prepare a short document from an approved set of sources. A weak definition of completion is that a file exists. A useful definition might require accurate citations, the requested structure, no unsupported claims and acceptance by the person responsible for publication.
Generate ten files and accept six of them under those criteria. There are six completed jobs in the denominator. The other four still used model capacity, elapsed time and someone’s attention. Keeping their costs in the total makes the comparison less flattering and more useful.
Choose a representative set of tasks and keep the acceptance rule stable. Record attempts, accepted outputs, review time, repairs and work handed to another process. A spreadsheet is enough to begin. Its value comes from preserving the awkward cases alongside the easy ones.
The broader idea of total cost of ownership supplies useful context: the visible purchase price is only one part of the overall cost. For AI work, the categories should match the actual job. A casual first draft, a customer-facing answer and a document supporting a consequential decision require different levels of review. Treating them as one generic task would hide the very costs the analysis is meant to reveal.
Follow the provider, the invoice and the working day
OpenAI's argument touches three related but distinct economic boundaries. The first is the provider's cost to serve an interaction. The second is the customer's price for access or usage. The third is the organization's total cost of obtaining an acceptable result.
A lower serving cost may produce a lower price, more capacity, a different tier or a better provider margin. The buyer has to observe which of those changes reaches the invoice. Then there is another question: did the work become easier to complete? A higher-priced model may still need fewer attempts and less correction.
The organizational ledger includes work outside the AI invoice. For the document example, that can include preparing the source set, explaining requirements, checking citations, correcting language and escalating exceptions. Some of those activities may also exist in the previous workflow, so the comparison must apply the same accounting rules to both paths.
Keep two kinds of time visible. Elapsed time follows the clock from request to completion. Human effort follows active attention. An unattended hour may consume less labor than ten supervised minutes, yet still miss a deadline. Separate columns let the team see both consequences before combining them into one cost.
Trace the hardware result into a particular job
The essay points to OpenAI’s August 25 Jalapeño inference-chip results , covering InferenceX tests on GPT-OSS 120B, DeepSeek R1 670B and Kimi K2.5 1T. OpenAI reports 1.5 to 1.9 times more work per watt at peak throughput and 1.7 to 3.6 times lower end-to-end latency. The comparison uses published accelerator power ratings for normalization. Its September strategy essay planned deployment to begin by year-end.
The hardware results describe the tested systems and workloads. A customer’s subscription price and electricity use sit elsewhere in the chain. To connect the measures, we would need evidence of what changed between the tested accelerator and that customer’s actual job.
For an agent, faster inference can shorten one step or make more capacity available. An accepted task may require several such steps, along with a check, a repair and an occasional fallback. Following that route shows where a hardware improvement reaches the workflow and where another constraint takes over.
A throughput number therefore belongs inside the cost comparison. It cannot stand in for the whole result. Apply the same boundary to any vendor or independent benchmark: preserve the workload it tested, then measure how much of the improvement survives the work the buyer needs done.
The cheaper attempt meets another retry
Return to the hypothetical document task. Imagine two systems judged against the same acceptance criteria. System A costs 20 pence per attempt and averages three attempts for an accepted result. System B costs 45 pence per attempt and averages one and a quarter attempts. Before human work, the expected model charge per completed task is 60 pence for A and about 56 pence for B.
That reverses the apparent price ranking before human work. Now suppose A’s corrections take two minutes and B’s take five. The labor comparison may reverse it again. A frequent document type that needs manual routing adds another cost. Split easy and difficult tasks where their behavior differs, so an average does not hide the exception that keeps consuming time.
Keep failed work in the record. A system that handles straightforward cases but stalls on a common exception needs a defined alternative. Detecting the failure, explaining it and moving the job to that alternative all take effort. The task has not become free because it left the AI route.
Our analysis of Nvidia's Personal AI Router examines a related distinction between routing a request and the infrastructure that serves the model. Here, the corresponding lesson is that an improved component should not stand in for the performance of the complete workflow.
The revealing row in the spreadsheet may be the rejected file. It has no accepted output beside it, but the review minutes and model charges are real. Include that row and the cost argument starts to describe the work a team actually has to finish.
