Gergely Orosz on token maxing and AI's impact on engineering roles
Gergely Orosz, author of The Pragmatic Engineer newsletter, discusses "token maxing" — the practice of artificially inflating AI token usage to game internal leaderboards at companies like Meta, Microsoft, and Salesforce.
Score breakdown
Engineering leaders and practitioners should scrutinize how AI usage metrics are tracked and communicated internally, as leaderboards and spend targets can incentivize performative rather than productive AI adoption.
- 01"Token maxing" refers to artificially inflating AI token usage to game internal leaderboards or meet spend targets at large tech companies.
- 02Companies named as having token-related tracking include Meta, Microsoft, and Salesforce.
- 03At Salesforce, Orosz describes a minimum monthly AI spend target of approximately $175.
In a conversation at AI Engineer, Gergely Orosz — former Uber and Skyscanner engineer and author of The Pragmatic Engineer, the #1 software/AI engineering newsletter on Substack — describes how "token maxing" has evolved from a playful early trend into a culturally problematic behavior at large tech companies. The practice involves engineers deliberately generating large volumes of AI tokens — not for productive work, but to appear active on internal usage leaderboards or to meet minimum spend targets. Orosz recounts hearing from engineers at Meta, Microsoft, Salesforce, and others, each with slightly different contexts but a common thread of anxiety: in an environment of industry-wide layoffs, workers fear that a low token count could be used against them in performance evaluations or promotion decisions.
At Meta, Orosz says, token count is one of several data points in performance reviews — alongside metrics like diffs, code reviews, and impact — and can be "weaponized" to flag low performers.
At Meta, Orosz says, token count is one of several data points in performance reviews — alongside metrics like diffs, code reviews, and impact — and can be "weaponized" to flag low performers. A leaderboard existed internally and was shut down after an article surfaced about it, though engineers reportedly continued token maxing anyway. At Salesforce, a minimum monthly AI spend target — which Orosz estimates at around $175 — has led engineers to front-load token usage at the start of each month. At Microsoft, he describes engineers running autonomous agents to produce what he characterizes as "junk" output, solely to inflate their numbers. Orosz frames this as a cautionary tale about how well-intentioned productivity metrics can distort behavior when tied to job security concerns.
Key facts
- 01"Token maxing" refers to artificially inflating AI token usage to game internal leaderboards or meet spend targets at large tech companies.
- 02Companies named as having token-related tracking include Meta, Microsoft, and Salesforce.
- 03At Salesforce, Orosz describes a minimum monthly AI spend target of approximately $175.
- 04Meta had an internal token usage leaderboard that was shut down after an article about it was published, but engineers reportedly continued the behavior.
- 05Token count is used as one of several data points in Meta performance reviews, alongside metrics like diffs, code reviews, and impact.
- 06Engineers have resorted to tactics like having agents summarize documentation they could read themselves, or running autonomous agents to generate low-value output.