Cheap tokens. Expensive outcomes.
Generation can become 90% cheaper while a completed outcome becomes only 4.3% cheaper. Follow cost through review, rework and the standard for acceptance.
Contents & sources
A faster AI answer accelerates one stage of production. Whether it has become a usable result is a separate question. Capital ultimately needs an outcome that a customer accepts and a system can use, rather than a volume of text.
What one experiment can tell us
METR’s 2025 experiment involved 16 experienced open-source developers completing 246 tasks in their own projects. AI access was randomised by task. In that setting, allowing AI increased completion time by 19%. The researchers did not claim the sample represented all developers or work.
Our question is not whether AI is slow. It is which task, acceptance standard and user experience determine which costs change. Producing a plausible answer and satisfying a particular organisation’s requirements are different capabilities.
Update the judgment when the evidence changes
In February 2026, METR described a redesign of its follow-up experiment. Developers and tasks were selected out when participants did not want to work without AI; concurrent agents also complicated time measurement. The team considered greater speedup plausible, but its magnitude weakly evidenced.
The 2025 result is not a fixed verdict on present systems. The follow-up also does not establish a universal productivity multiplier. Keep dates and conditions attached to the evidence.
Costs that survive the generation step
Our thought experiment reduces average generation cost per completed result from $0.10 to $0.01. Assume two minutes of review and rework in each case, valued at $60 per hour. Total cost falls from $2.10 to $2.01: a 90% reduction in generation cost produces about a 4.3% reduction overall. Change the assumptions below.
This is arithmetic under assumptions, not a company measurement or a return forecast. Faster review can increase the gain; more rework can reverse it. Review and rework are production stages to measure, not inconvenient extras to discard.
Creation and capture are different measures
Product prices can fall faster than production costs as competition expands. Track whether savings reach customer prices, compensation or supplier margins. Productivity created and profit captured are distinct quantities.
Our proposed measures are total cost per verified outcome, acceptance rates, human time, rework and delivery costs. Inspect those before deciding which productive system is worth owning.
Follow the price of an answer all the way to the cost of an accepted outcome.
Primary sources checked
An analysis linking Kamui’s 19 September 2026 post with observed productivity research, checked on 4 October 2026. The calculator performs arithmetic under selected assumptions. It is not observed data, a benchmark or an investment-return forecast.