In March 2026, The New York Times reported on tokenmaxxing, a practice in which engineers competed on internal leaderboards to consume the most AI tokens because they believed high usage signaled a serious commitment to AI. By midyear, those leaderboards had disappeared, and public model benchmarks were reporting cost per task instead. This shift represented genuine progress, as cost per task became the standard unit. Tokens are what providers bill for, but tasks are what businesses need completed.
Still, there's a problem with treating cost per task as the final measure of AI efficiency. An agent can finish a task, return a convincing answer, and get the answer completely wrong.
If that happens, what did the task actually cost? And if the answer depends on how an agent accesses and interprets enterprise data, are we measuring the right thing when we compare models alone?
Three measurements in nine months
AI cost measurement has gone through three distinct phases in a remarkably short time.
The first phase was tokenmaxxing, based on the idea that consuming more AI tokens demonstrated greater adoption, experimentation, or productivity. But token volume only tells us how much AI was used, not whether that usage produced anything valuable.
Then came a sharper focus on tokenomics, the economics of token-based AI consumption. Teams started examining model pricing, caching, prompt efficiency, and ways to reduce unnecessary token usage. Deloitte now frames tokens as the currency that translates infrastructure decisions into economic terms, which is a fair description of how most teams budget AI today.
The next logical step was measuring the work itself. Call it taskonomics: evaluating AI economics in terms of completed tasks rather than tokens consumed.
In June 2026, Artificial Analysis introduced cost per task, time per task, and tokens per task into its Intelligence Index. Its subsequent benchmarking made it easier to compare model capabilities against their actual task costs.
Each metric improved on the last. A model with a low per-token price may require more reasoning, retries, and tool calls to complete a job, while a pricier model may finish it at a lower total cost. Cost per task endured because it was the first metric to measure completed work.
Why the denominator is the hard part
Cost per correct task is the total cost of a workload divided by the number of tasks it completed correctly, not the number it attempted or returned. Unlike cost per task, this calculation excludes responses that finished but failed to produce the verified result.
In CData Labs' 22-model benchmark, our researchers tested a task requiring models to combine two business signals across live CRM, warehouse, and ITSM data: find the high-priority accounts that are both in poor health and declining in usage.
With raw access to all three systems, not one model was able to produce the correct account set. The best score was 0.45. Eighteen of the 22 models scored zero.
These weren't simply incomplete attempts or obvious system failures. The models returned well-formed account lists that looked like legitimate answers, with nothing anywhere in the stack flagging that the composition logic had been misapplied. A wrong answer that looks exactly like a right answer still registers as a completed task. This demonstrates that cost per task, measured on its own, can report a perfectly healthy number on a workload that is failing.
When context carries the business logic
The benchmark revealed something more important than which model performed best: the way an AI agent accesses business data can have as much impact on accuracy as the model itself.
When models were given raw access to enterprise systems, they had to reconstruct the business logic needed to complete the task. Most got it wrong. But when that logic was defined once in curated tools and made available through the same gateway, models across different price tiers produced the same correct result.
That changes how organizations should think about AI cost optimization. Rather than relying on increasingly capable—and potentially more expensive—models to interpret complex enterprise data, teams can move repeatable business logic into the data access layer. This makes results more consistent while giving organizations greater flexibility to choose models based on cost and performance.
In fact, the CData Labs benchmark found a 178x cost difference between models producing the same correct answer.
The straightforward implication is that AI cost efficiency isn't just a model-selection problem. It's also a data architecture problem. When correctness is built into the context an agent receives, organizations have more freedom to optimize the latest measurement in AI spending: cost per correct task.
Run your own taskonomics
The spread is easier to believe once you have watched it on your own systems. Connect a source, run the same question through a cheap model and an expensive one, and compare what each correct answer cost you.
Start a free trial of CData Connect AI today!
Frequently asked questions
What is cost per task in AI?
Cost per task is the total spend for a workload divided by the number of tasks it completed, rather than by the number of tokens it consumed. Public model benchmarks began reporting it as a standard metric in 2026. It is a better budgeting unit than token price because a cheaper per-token model can take more steps and cost more per finished job.
Is cost per task the same as cost per correct task?
No. Cost per task counts every task that finished, including the ones that returned a confident wrong answer. Cost per correct task divides by the tasks verified right against a known answer, which is the number that reflects what a workload actually produced.
Does switching to a cheaper model reduce AI accuracy?
It depends on where the business logic lives. On raw data access the model is deriving that logic itself, and model tier matters. When the gateway carries the logic, a 22-model benchmark found accuracy equalized across tiers while cost still spanned 178x for identical output.
