Methodology
Where every number comes from
The Token Tape is a measurement terminal, not a review site. Every figure is fetched from a named public source on a fixed schedule, stored with its source URL and raw payload, and ranked by the same deterministic formula for every vendor. Nothing on this site is an opinion score.
Data sources
| Source | What we take | Cadence (UTC) | Notes |
|---|---|---|---|
| OpenRouter datasets API | Daily app and model token rankings (usage leaderboards) | daily · 06:00 | Authenticated API. Republishing these numbers requires the citation “Source: OpenRouter (openrouter.ai/apps or /rankings), as of {date}” — our API passes that note through. |
| OpenRouter models API | Model catalog, per-MTok prices, context windows, deprecation dates | daily · 05:30 | Public endpoint; prices cross-checked against the two catalogs below. |
| llm-prices.com | Current + historical per-MTok prices | daily · 05:30 | Community-maintained. Future-dated (announced) prices are stored but never shown as in effect. |
| models.dev | Open-source model catalog: prices, limits, release dates | daily · 05:30 | Third price source; disagreements surface as separate rows, each with its own source URL. |
| Vendor pricing pages | Subscription plans, tiers, limits for each tool | daily · 07:00 | Rendered to text, snapshotted, and machine-diffed; a plan change is only reported when two dated captures differ. Client-side price toggles some vendors use can hide non-default tiers — we only diff what was actually captured. |
| Changelogs & GitHub releases | Release notes and versions for tools and models | daily · 07:00 | Straight from each vendor’s changelog or GitHub releases feed. |
| GitHub API | Stars, forks, open issues per tracked repo | daily · 06:30 | Daily snapshots plus starred-at deltas via GraphQL. |
| npm & PyPI registries | Daily package download counts | daily · 06:30 | Registries occasionally report false zero-days on high-volume packages; we flag those as gaps instead of charting them as drops. |
| Mention counts and sentiment for each entity | daily · 07:00 | Official API. Mention counts are exact; sentiment labels are language-model estimates and are presented as such. | |
| SWE-bench experiments | Published SWE-bench Verified / Lite results | daily · 07:00 | We ingest published results only — we do not run evaluations ourselves. |
| Aider polyglot leaderboard | Polyglot coding benchmark results | daily · 07:00 | Ingested from the Aider repo’s published data file. |
| Wayback Machine | Historical backfills of public rankings and pricing pages | one-time backfills | Used to extend history before our tracking began; rows keep the archive snapshot URL. |
every stored datum keeps its source url + raw payload
pipeline last reported 42 min ago
Ranking formula
Deterministic, and the same for everyone
Leaderboard rank is a descending sort on total tokens (prompt + completion) from the latest daily OpenRouter snapshot for the selected category — all, coding, or CLI agents. No weighting, no editorial adjustment, no exceptions.
Deltas compare the latest snapshot to the closest stored snapshot at least 7 days older (or the window you select). If no older snapshot exists, we show no delta rather than an estimated one.
Crossovers (“X overtakes Y”) use 7-day token sums to smooth daily noise: the window ending today (W1) versus the window ending yesterday (W0), across the top 25 apps in the coding category. A crossover is only reported when the pair’s order actually flipped between W0 and W1 and the new leader’s 7-day margin is at least 5% — so a lead that flaps on noise never makes the tape.
Known limits: OpenRouter traffic is a large, public, consistently-measured slice of the market — but it is not the whole market. Direct-API and subscription usage is invisible to it, and token counts come from each provider’s own tokenizer, so cross-provider comparisons are directional, not exact. We repeat this caveat wherever the numbers appear, including in the API’s source_notes.
Update cadence
| UTC | Run |
|---|---|
| 05:30 | model prices refresh (OpenRouter models API, llm-prices.com, models.dev) |
| 06:00 | OpenRouter usage datasets → leaderboard snapshots |
| 06:30 | adoption: GitHub repo stats, npm/PyPI downloads |
| 07:00 | changelogs, vendor plan pages, benchmarks, Reddit pulse |
| 07:30 | event detection sweep (price/plan/rank/release/benchmark diffs) |
| 08:00 | daily brief published |
| hourly :15 | email alerts for new significant events |
| Mon 09:00 | weekly digest |
pages rebuild at most hourly (isr 3600) and on demand after data runs · all date logic utc
Independence
No vendor pays for rank
Rankings are a sort over measured public data. There are no sponsored placements, no affiliate reordering, and no paid inclusion — a vendor cannot buy a position on any table here, and we don’t accept payment to add, remove, or reclassify an entity. If we ever sell anything (like the Pro data plan), it buys deeper access to the same numbers everyone sees — never different numbers.
Where a source itself has a conflict (a vendor’s own pricing page, a benchmark run by a model’s creator), the row says so via its source link, and subject-reported results sit next to independent ones rather than replacing them.
How this site is operated
Agent-operated, human-signed
The Token Tape is run by an autonomous data pipeline: scheduled jobs fetch the sources above, diff new captures against history, detect events (price changes, plan changes, rank crossovers, releases, benchmark results), and publish the daily brief — without a human in the loop. Machine-written copy is generated only from detected, stored events, so it can only cite numbers that exist in the dataset with a source attached.
Humans set the methodology on this page, review it, and answer for it. The formula changes only when this page changes — and this page is versioned in the site’s repository.
methodology signed by: Ryan Doyle · operated autonomously since 2026
Corrections
Found a wrong number?
Email ryan@thetokentape.com with the page URL and a link to the correct source. Because every datum keeps its raw payload, we can audit any figure back to the byte we ingested. Verified corrections are applied to the stored data — not silently patched onto a page — and affected pages rebuild within the hour.
Reusing this data
Everything here is free to republish under CC BY 4.0 with a link to thetokentape.com. The free JSON API serves the leaderboard, entities, and the event tape; the RSS feed carries briefs and significant events. OpenRouter-derived numbers additionally require OpenRouter’s own citation, which the API includes in meta.source_notes.
data as of Jul 8, 2026 · updated daily · thetokentape.com