methodology

Tiers + Teams

Your all-time tokens put you in a tier. Your model split puts you on a team. That's it. No composite score, no sigmoid formula, no five dimensions. Just your /usage, ranked against everyone who showed up.

How tiers work

Your tier is a percentile against the active cohort (users who have real token activity on file). One metric drives it: all-time tokens, counted as input + output + cache_creation + cache_read (the ccusage formula). No weights, no adjustments, no gaming the curve.

Percentile is computed fresh against the active cohort on each load, so early-field noise never locks you into a wrong tier. The bands:

TIER

BAND

WHAT IT MEANS

Top 1%

Lets it cook. One-shots SaaS apps at 3am and reviews the PRs while you sip coffee. The ceiling everyone else is measured against.

Top 10%

Dispatches agents in parallel worktrees and burns the rate limit by noon. One tier off cracking S and you know it.

Top 40%

Two-shots it on a good day. Survived context rot, hit /compact, lost the plot, shipped anyway. Respectable.

Top 75%

Opens the agent when stuck, closes it when nervous. The come-up is right there. You just keep closing the tab.

75-90%

Showed up to the casino, bought one chip. One bad week from the handcoder label. Pick a team and actually play.

HANDCODER

Bottom 10%

Bottom 10% of the active field, or zero AI tokens on file. Two sub-cases: if you have zero tokens, there's nothing to rank. If you have tokens but land in the bottom 10%, you out-purist Linus Torvalds, and he caved. Pick a team and climb.

The token count

all_time_tokens = SUM(
  input_tokens
  + output_tokens
  + cache_creation_tokens
  + cache_read_tokens
)

This is the ccusage formula, applied across every session ever ingested. Cache reads count because they represent real context load, even if the tokens were created in a prior session. All-time volume is the right signal for a leaderboard: it captures who has been in the trenches the longest, not just who had the biggest single day.

Team Claude Code vs Team Codex

At signup you pick a camp. The split is derived from your token data:

Team Codex: tokens from gpt-* models (OpenAI's Codex and GPT family). If your tokens_by_model breakdown shows a heavy gpt-* footprint, you belong here.
Team Claude Code: everything else. Sonnet, Opus, Haiku, any non-gpt-* model routed through Claude Code.

A live daily Team Scoreboard sums each camp's tokens across the entire field and shows the running split. It swings. Check it daily or don't, but your team's position is never static.

You get one free switch per month (rolling 30 days). A switch earns you a DEFECTOR badge for a window. Bold move. We saw that.

What drives the card

Your OG card shows: all-time tokens (the hero number), your tier letter, your global rank, your percentile, your team chip, and supporting stats (peak day, sessions, active days). The gap to your next tier is visible on the card so you know exactly what it takes to climb.

One number, one letter. Your /usage says everything. So does your tier.

What this is not

Not a measure of output quality. Not a claim that more tokens means better work. It's a tokenmaxxing leaderboard: you showed up, you burned tokens, here is where you rank against everyone else who did the same. The flex is the number; the tier is the context.

Holden is #1 global at 4.84B all-time tokens, S-tier, Team Claude Code. That's the anchor the top 1% is measured against.

Questions, calibration feels off, found a gaming exploit? Open an issue at github.com/holden-alt/vibecodestats. <- back to dashboard