How this is measured

A tracker that will not say what it does not measure is not worth trusting.

board captured 3 August 2026 · 36 tools watched · 339 models eligible · read once a night

Last run finished

3 August 2026

2026-08-03Freshness is the whole claim, so it is also the health check.

Tools watched

2026-08-03Including the ones that were acquired, archived or sunset.

Models eligible

339

2026-08-03Everything that fails the eligibility rule is excluded before ranking.

Changes in 7 days

2026-08-035 of them high severity.

The board

Best coding model right now

It measures: Adoption — how much a model is actually run inside the tracked coding tools.
How the rank is computed: A weighted score, not a raw volume sort: 65% token volume inside the tracked coding tools, 25% overall OpenRouter volume (a stability term — per-tool data is thin and a model that is large everywhere should not be invisible), 10% breadth across tools, capped at four. Each is normalised against the leader in that band, so the Volume column can disagree with the rank order and usually does. The Volume column on every board is one of those three inputs, printed raw — so a row can sit above another with less volume next to it. That is the function working, not a sorting bug.
It does NOT measure: Capability. No benchmark is involved. A cheap model can lead on volume alone. This is the most important line on the site. A model at rank 1 is the one being used the most, not the one that writes the best code, and nothing here is evidence of the second thing.
The coding signal: Usage inside tools that are coding tools, summed per model. OpenRouter retired its per-category programming ranking, so category tags are no longer available; measuring the tools directly replaces it.
Price basis: Blended USD per million tokens, weighting prompt to completion 3:1, because coding traffic is prompt-heavy. Where a model appears in usage data with no published price, the board says price not published — never zero. A free model and an unpriced one are different facts.
Which models qualify: Text input, tool-calling support, and at least a 32K context window.
The per-tool boards: Where a tool bundles a model, that model is listed first because it is what a user gets by default — it is not scored against the rows under it. Measured rows below it are rank-ordered by that tool's own token share and reflect only users who supplied their own OpenRouter key.
Movement: A rank is compared against the previous night’s capture of the same board and marked as having risen, fallen, held or entered. Today's board is measured against 2 August 2026. A missed night means no comparison rather than a comparison against a stale capture.

Two kinds of fact, never dressed alike

Every figure on this site is one of two things, and the difference is structural rather than a footnote. A number this site counted — a rank, a price band, a token volume, a star count — is set in the display face, in the accent, beside the day it was counted. A claim a vendor published about its own product — a context window, a licence, a status note — is set in Inter and never carries the accent. A reader who scrolls past a table without noticing which is which has been failed by the page.

Measured — counted from real repositories, nightlyAsserted — read from the vendor docs, dated

The change feed

What counts as a change

What is read: Each tracked tool's own pricing page and its home page, plus its public repository metadata, fetched once a night and compared against the previous fetch. A difference in what a tier costs, which tiers exist, what a quota allows, or who owns the thing is a change.
Why the home page too: Because that is where a company writes that it has been acquired or that it is shutting down. Until August 2026 this pipeline read only pricing pages and repositories, and it missed two of them sitting in plain text on the vendors' own front pages. A phrase like “was acquired by” appearing where it was not before is recorded as a signal to check, at low confidence and with the link — it never flips the status column by itself, because “acquired” carries a who and a regex cannot be trusted with the who.
What is not a change: A feature shipping. A blog post. A redesign of the pricing page that leaves the prices alone. This is not a release-notes mirror and it does not try to be one — if it does not alter what the tool costs, what you get free, or whether it still exists, it does not appear.
How sure each entry is: Entries derived from an API — a repository archived, an ownership record — are marked high confidence. Entries read off a marketing page are marked medium or low, and the row says so rather than presenting everything as equally certain. Every entry carries a link to the evidence.
Withdrawn is not shut down: A tool that ends a free tier has not stopped existing. Gemini CLI ending free consumer access in June 2026 is recorded as a withdrawn tier, because the tool still runs on a paid key. Collapsing the two is how a tracker ends up wrong in a way people repeat.

What this cannot tell you

The limits, stated

Which model is better: There is no benchmark here at all. Volume measures what people reach for, which is shaped by price, defaults and habit at least as much as by quality.
All of a tool's traffic: Usage is measured where it passes through a router that publishes it. Traffic that goes straight from a tool to a model provider is invisible here, and for tools that bundle a model that is most of it.
Enterprise pricing: Only prices published on a public page are read. Anything behind a “contact us” is not a number this site has.
Today: Every figure is from the last completed nightly run. A price that changed this morning is a price this site does not know about yet.
Plan pairings read from a page: Where a price is marked as scraped, the number is real but the tier it was matched to was inferred from its position on that page. The evidence link sits next to it for exactly this reason.

Check it yourself

Everything on this site is rendered from a public JSON API with no authentication and open CORS, so any claim here can be checked against the data it came from — and used in your own work without asking. No key, no rate limit. ToolDrift is free to read and has no paid tier.

Open data, no key

/api/leaderboard: Every board, grouped and rank-ordered, with this method object attached
/api/tools?include=plans: Every tracked tool with its plans
/api/changelog: The change feed with evidence links
/api/rankings: The raw router usage the board’s opinion is built on
/api/stats: Coverage and the age of the last run

Back to the board The change feed