Is a 3x AI Productivity Boost Just a Computer That Never Sleeps? (3 min read)

OpenAI revealed that leveraging AI agents achieved productivity equivalent to 3.14 agent-workdays per 8 human work hours. This is attributed to researchers running an average of four agents in parallel, effectively allowing AI to operate 24 hours a day.

However, this productivity boost comes with substantial costs. The median inference cost for OpenAI researchers skyrocketed 40x in five months, rising from $14 per day at the end of March to over $600 per day by mid-August. The top 90% of researchers consume over $7,000 per day, amounting to $2.5 million annually.

The Paradox of Cost Structure and Productivity

AI inference costs are classified as pure Operating Expenses (OPEX), with no depreciation of physical equipment. Unlike an automobile factory recovering capital expenditures (CAPEX) through night shifts, AI can operate at night with only variable costs. However, the figure of 3.14 days is not the result of 'smarter thinking,' but merely the outcome of machines running three shifts while humans are awake.

High Defect Rates and the Need for Human Intervention

The digital assembly line suffers from high defect rates. Over the past six months, more than half of successful tasks that took 4–8 hours required human intervention. OpenAI acknowledged that these metrics may not reflect the overall pace of progress.

• Cost Efficiency: While computing expenses increased 40x, work time only tripled.
• Role Shift: Engineers' roles have degraded from creative architecture design to 'factory floor management,' fixing machine errors that occurred overnight.
• Autonomy Limits: Only 50% of the 16-hour (two night shifts) machine operations are completed autonomously, resulting in an actual output of only about 2x.

Ultimately, the '3x productivity leap' expected by the market resembles 'paying for 2nd and 3rd shift labor' from a CFO's perspective. Currently, AI demands high costs and low autonomy as the price for being a machine that never sleeps, and quiet dissatisfaction is spreading among software engineering practitioners.