Tokenmaxxing is dead, long live tokenmaxxing

The tokenmaxxing story started as a punchline. Meta reportedly tied performance reviews to token usage, and predictably, people gamed it — one employee described having two agents just chat with each other all day to pad the numbers. It got mocked as executive incompetence, a bunch of leaders accidentally torching money on nothing.

But that framing misses what was actually happening. A few months back plenty of senior, respected employees simply refused to touch AI tools, and no amount of gentle persuasion moved them. Blunt policies that forced token spend weren't an accident — they were a lever to break through stubborn resistance. And it worked: nearly everyone now uses something like Cursor at least occasionally. The unfortunate timing is that OpenAI and Anthropic, both circling IPOs, chose this exact moment to tighten subscription limits and hike API prices, killing the subsidies that made unlimited spend painless. So companies are rolling back their token-maxing mandates, and the trend looks dead.

Except it probably isn't, for a more interesting reason than nostalgia. Until recently, letting an agent run unsupervised for hours meant watching small errors snowball into what practitioners call compounding error — mistakes that get baked permanently into a codebase. That capped how much token spend actually helped. Now the dynamic has flipped into what's being called compounding correctness: more tokens genuinely produce better outcomes, especially in long-running agent loops. Anthropic's unreleased security model, Mythos, is the starkest example — given 100 million token budgets per attempt, it kept finding new exploits with zero sign of diminishing returns, turning cybersecurity into a grim economic contest where whoever spends more tokens wins.

That same logic is fueling excitement over agent 'loops' — essentially the old Ralph Wiggum trick of just re-running a prompt until a large task gets chipped away, no supervision needed. It's not new, but it finally works reliably because each iteration tends to improve rather than degrade. This favors open models more than people expect: if a cheaper model like GLM 5.2 improves 5% per loop versus Claude's 10%, but costs a fifth as much, running it five times over still wins on cost and results.

A separate, uglier flavor of tokenmaxxing involves companies hiring consultants to build bespoke 'agents' for tasks that should've been plain deterministic code — data labeling pipelines, one-off automations — which then burn tokens without ever matching the reliability of normal software. Layer on 'quality-checking' agents to patch the errors and costs triple for no real gain. That version of tokenmaxxing deserves to die. But the underlying incentive to spend on tokens hasn't disappeared, it's just moved toward generalist platforms and always-on 'software factories' that write, test, and fix code with no humans in the loop. StrongDM's suggestion that engineers should burn $1,000 a day in tokens sounds absurd today, but the direction is not.