Metrics

Where the numbers come from

What the tokens are

A token is a small piece of text, roughly three or four characters, that models read and write. Claude Code counts four kinds.

Totals (“15.8B tokens”, “79% of all”) add all four kinds, as /stats does. Levels use output only.

Levels

Level = 1 + the number of targets the model's all-time output has reached.
Targets: 1K, 2K, 5K, 10K, 20K, 50K, 100K … (1, 2, 5 in every power of ten).

// Sources/ClepsoApp/Views/Details/ModelsSection.swift
static let targets: [Int64] = {
    var all: [Int64] = []
    var decade: Int64 = 1_000
    while decade <= 1_000_000_000_000_000 {
        all += [decade, decade * 2, decade * 5]
        decade *= 10
    }
    return all
}()

init(output: Int64) {
    let output = max(0, output)
    let reached = Self.targets.prefix { $0 <= output }.count
    self.output = output
    self.level = reached + 1
    self.floor = reached == 0 ? 0 : Self.targets[reached - 1]
    self.next = reached < Self.targets.count ? Self.targets[reached] : Self.targets[reached - 1]

Why output: it is what the model produced. Total tokens are dominated by cache reads, so a long conversation would level a model up without it doing more. Why fixed 1-2-5 targets: every next target is a round number you can read, and each level asks for 2 to 2.5 times the output of the one before, so levels keep meaning something at any scale.

Worked example: why “18.2M of 20M output” is progress to LV 15

A model has written 18.2M output tokens. The highest target reached is 10M, the 13th target, so it is LV 14. The next target is 20M, which makes LV 15. The level bar in its dossier runs from 10M to 20M: (18.2M − 10M) ÷ (20M − 10M) = 82%, with 1.8M to go.

The ladder

The ladder continues the same way past LV 25. In the app, “How levels work” under Stats › Models (Arena) draws it as a pixel track: a cell per level, at least 1 to 20, each target under its cell, every model's character standing on its level and the chosen model's next level marked NEXT. On 30-day and 7-day ranges, levels still use all-time output, because Claude Code records tokens by kind for all time only.

Rank and specialty

Rank: by level, then by output, then by total tokens. A higher rank never has a lower level.

Specialty ratio for a kind = (model's tokens of that kind ÷ model's total) ÷ (all models' tokens of that kind ÷ all models' total)

The specialty is the kind with the highest ratio, shown as, for example, “Reading input · 13× the party”: that model's share of input is 13 times the share across all models. If no kind reaches 1.1×, the model is an All-rounder. Every model mostly reads from cache, so the plain largest kind would say the same thing on every card; the ratio shows how each one differs.

Agents per task

Optional, off until you turn it on in Settings. A task is one prompt: from the moment you send it until Claude stops. Clepso adds four background hooks to Claude Code (prompt submitted, subagent started, subagent stopped, Claude stopped) and counts the subagents each task starts, nested ones included.

Limits and the pace estimate

A state clears only after usage drops 3 points below its threshold, so it does not flicker.

// Sources/UsageEngine/PoolReducer.swift
static func classify(_ used: Decimal, previous: LimitState, config: EngineConfig) -> LimitState {
    if used >= config.exhausted { return .exhausted }
    if used >= config.critical { return .approaching(.critical) }
    if previous == .approaching(.critical), used > config.critical - config.hysteresis { return .approaching(.critical) }
    if used >= config.warning { return .approaching(.warning) }
    if case .approaching = previous, used > config.warning - config.hysteresis { return .approaching(.warning) }
    return .available

Pace estimate (always labelled “Estimate”): a straight line fitted by least squares through recent readings in the current window: the last 90 minutes for the 5-hour limit, the last 24 hours for the 7-day limit. It needs at least 3 readings over at least 10 minutes. At or below 0.5 points per hour usage counts as flat (“not rising”). It never looks further ahead than 8 times the history it rests on, so an hour of readings cannot predict two days. If the window resets before the line reaches 100%, it says roughly how much will be used by the reset.

// Sources/UsageEngine/Pace.swift
guard slope > flatSlope else { return .notRising }

let hoursToLimit = (100 - latest.usedPct) / slope
let hit = latest.at.addingTimeInterval(hoursToLimit * 3600)
let span = latest.at.timeIntervalSince(first.at)
let horizon = min(hit, resetsAt.flatMap { $0 > now ? $0 : nil } ?? hit).timeIntervalSince(latest.at)
guard horizon <= span * maxLookAhead else { return .notEnoughData }
if let resetsAt, resetsAt > now, hit >= resetsAt {
    let projected = latest.usedPct + slope * resetsAt.timeIntervalSince(latest.at) / 3600
    return .belowAtReset(projectedPct: min(projected, 100))
}
return .reachesLimit(at: hit)

The Limits chart spans the limit's own window, from reset minus its length to the reset. The hatched span before the first reading is time Clepso was not yet watching.

Session figures

Activity calendar

Each square is a day. Its level is set by messages that day, in quarters of your active days, and drawn as pixel density rather than shade: the quietest quarter lights a few pixels, then a checkerboard, then solid with a few holes, and the busiest quarter is solid. The pattern says the level without relying on colour. A day with no messages is the plain square; a day before your first session or after today is an outline marked “no data”. All time shows one calendar year (January to December) with a year switcher that goes back to 2025; 30 days shows a month calendar, each day's date and message count on the plain square with its level as a pixel band below them; 7 days shows a column per day.

Charts

Tokens per day are stacked by model on a real date axis: days with nothing stay on the axis. Past 120 days the bars become weeks. Every bar chart is drawn as pixel columns: whole-pixel widths, heights in whole pixels from the baseline, stepped top corners. The scale's top is the first round number at or above the tallest bar whose quarters are also round (for example 8M for a 6.5M peak). Models differ by colour and fill pattern, never colour alone. Point at a bar for its numbers.