Why 'tokenmaxxing' isn't dead — it's just getting a smarter incentive

Tokenmaxxing—companies pushing employees to burn through AI tokens, most infamously Meta tying performance reviews to per-person token usage—predictably led to gaming, including engineers leaving two agents chatting all day to pad their numbers. The author’s contrarian read is that this wasn’t executive stupidity but deliberate blunt force: a way to drag AI-resistant senior staff into actually using the tools. By that measure the policy worked, since nearly everyone now codes with AI assistance at least occasionally.

The catch is timing. Rising token spend collided with OpenAI and Anthropic preparing to go public, both of which throttled subscription limits and raised API prices as subsidies dried up. With incentives gone and costs up, teams are scrapping unlimited-spend policies—hence ‘tokenmaxxing is dead.’ But the author argues a stronger incentive is replacing the artificial one: a shift from ‘compounding error,’ where long unsupervised agent runs drifted into unrecoverable mistakes, to ‘compounding correctness,’ where spending more tokens reliably yields better results. That flips the economics back toward heavy token use.

The security angle makes the point concrete. The piece cites Anthropic’s ‘Mythos,’ an unreleased model so capable at security tasks that access was limited to critical software vendors so they could harden systems first. With AISI budgeting 100M tokens (~$12,500) per attempt and seeing no diminishing returns, security starts to look like a proof-of-work contest: whoever spends more tokens hunting exploits wins. The same logic fuels enthusiasm for agent ‘loops’—restarting a prompt until a large spec is chipped away autonomously, an idea previously dubbed the ‘Ralph Wiggum loop.’