Anthropic Report: AI Agents' Work Hours Reach 6 Times That of Humans, Share of Junior Tech Job Openings Falls to 9%

Anthropic report: AI work hours already 6 times those of humans, entry-level programmer positions plummet, revenue per employee is 2 to 4 times that of Meta and Google at the same stage

English text is being prepared.

English text is being prepared.

English text is being prepared.

This figure is based on Anthropic's internal financial and headcount records as of April 2026, and is compared, after inflation adjustment, with publicly reported financial data from Meta and Google five years after their founding.

The report shows that the share of entry-level employees in Anthropic's technical departments has fallen from 21% in early 2025 to 14%, and the share of entry-level technical positions in hiring has plunged from 27% to 9%. About one-third of machine learning managers and one-fifth of software engineering managers said they have reduced hiring because of the emergence of AI. At the same time, the company as a whole is still expanding, with positions such as sales, finance and recruiting growing faster.

Agent work hours exceed human work hours. Note: We consider only employee-led interactive sessions and the sub-agent pool. The concept of work hours refers to agent runtime, including reasoning and tool use, but excluding waiting times of more than 30 minutes. Human work hours are calculated by headcount, assuming 40 hours per week.

AI has brought not only job contraction but also a leap in efficiency. After adopting coding agents, the average number of code commits per engineer reached 5 times the level of 6 months earlier; lawyers handle more contracts, finance handles more invoices, and salespeople close more deals. After tracking 192 engineers who adopted coding agents, the report found that their code output rose markedly after adoption, and the proportion of code later deleted or rolled back did not worsen.

English text is being prepared.

Productivity surged across functional departments. Note: A Loess smoothing method was used to fit the underlying indexed output data.

But "human-AI collaboration" also brings new burdens. The report says that even for non-technical employees, reading all the output of their Claude agents could take up half a workweek; employees are shifting from "executors" to "agent managers," responsible for writing instructions, building monitoring dashboards and reviewing AI output. In spring 2026, this kind of "meta-work" already accounted for 22% of employees' interactive Claude sessions, up from 10% the previous year.

Estimates of productivity gains based on individuals' gradual adoption of AI. Note: The coefficients compare users who adopted coding agents with non-adopters, and account for two-way fixed effects for users and calendar months. Treatment time is based on the month in which a user first had at least 5 agent sessions. The bars represent 95% confidence intervals based on standard errors clustered by user. The quality panel compares code written directly by users with code written by the same users using coding agents. Quality metrics include: the share of code lines deleted within 90 days for any reason; the share of code lines deleted specifically for bug fixes; whether an entire commit was explicitly rolled back; whether code-check suppression comments were added; whether comments were added that require fixing; and whether tests were marked as skipped, retried or allowed to fail.

Hiring standards have changed accordingly: managers place more value on candidates' autonomy, judgment and ability to use AI creatively, rather than narrow domain expertise and specific execution skills. Anthropic has also established a job-risk early-warning mechanism and prioritizes internal reassignment of affected employees.

Entry-level employment growth by Claude usage complexity. Note: The chart is by job position. The growth rate compares employment in January 2026 and July 2026. AI complexity is based on automatic coding of agent logs (before the decline in employment, i.e., October to December 2025), and was percentile-transformed. The dots in the chart are scaled by the position's employment in January; the regression line is fitted to the weighted points.

However, revenue per employee is not the only indicator of corporate health. External research institute Epoch AI previously estimated that Anthropic's annualized revenue per employee in April 2026 was about $9 million, significantly higher than Meta's about $2.55 million and Google's about $2.11 million; but the institute also cautioned that its employee count and revenue data were not disclosed for the same period, so the estimates carry uncertainty.