Token leaderboards are apparently a thing now.
A developer shared that their company handed every employee unlimited access to Claudee Code](/tools/claude-code/) - Anthropic's AI coding assistant that runs inside your terminal and editor - and then added one twist: a public weekly leaderboard tracking who consumes the most tokens. Tokens are the units AI models use to process text and code; a rough rule of thumb is about 750 words per 1,000 tokens.
The competitive layer is clever. Most companies rolling out AI coding tools run into the same problem: they buy the licenses, adoption stays shallow. Developers use AI for the occasional autocomplete and move on. Tying usage to a public score changes the incentive structure entirely.
What Actually Drives Token Consumption
For anyone trying to maximize usage (or just get more value out of Claude Code), the patterns that burn tokens most effectively are consistent: longer iterative coding sessions rather than quick one-off queries; feeding in entire files for review rather than snippets; using Claude for tests, documentation, and refactors alongside net-new code; and staying in dialogue with the model through multi-step debugging rather than copy-pasting results into a separate editor.
The heavier the back-and-forth, the more tokens burn. The leaderboard effectively rewards usage patterns that tend to produce the best outcomes - sustained, conversational coding sessions rather than shallow dips.
The Goodhart Problem
There is a real argument that this rollout approach beats most corporate AI initiatives. Instead of a slide deck about AI productivity and a wiki page of prompting tips, you get visible social proof of who's actually using the tool and how hard.
The risk is that people optimize for token count rather than output quality - spinning up elaborate prompts that burn tokens without producing better code. When a measure becomes a target, it stops being a good measure.
Whether the leaderboard is tracking the right thing matters less than the adoption effect. Companies that make AI tool usage visible tend to get more of it. And more usage, even imperfect usage, tends to produce faster skill development across the team than any amount of internal training materials.