Methodology / TOKENRANK
TokenRank data methodology
TokenRank is T Salon's own leaderboard of AI coding token consumption. This page documents where the data comes from, how it is counted, and where it should not be relied on.
Where the data comes from
Data is submitted by developers who voluntarily install the T Salon reporting agent. The agent tallies token usage from AI coding tools locally, buckets the totals by tool and model, and uploads the aggregated figures. Developers who never install the agent contribute nothing.
Three counting bases
| Basis | Definition | Question it answers |
|---|---|---|
| Total | All token consumption including cache hits, reflecting real request volume | How much does this developer call |
| Normalized | Net token consumption after removing cache reads | How much computation actually happened |
| Estimated cost | Dollar amount derived from model pricing | What is this usage roughly worth |
Time bucketing
All data is bucketed by Beijing time (UTC+8). Daily usage is the delta between the current cumulative value and that device's previous snapshot, so a day only has a record if the device uploaded data for it.
Identity merging
A single developer may run the agent on several machines. TokenRank identifies devices belonging to the same person by GitHub account ID and merges them, so each leaderboard row represents a person rather than a device.
Cost estimation
Cost estimates use a model pricing snapshot dated 2026-08-26, sourced from theofficial OpenAI pricing page. Usage that cannot be matched to a specific model falls back to a default tier, so cost is an estimate rather than a billed amount.
Known limitations
- Self-selected sample: the data comes from developers who chose to install the agent and does not represent all developers.
- Cost is an estimate that depends on subscription type, discounts and billing arrangements; it should not be treated as a billed amount.
- Historical figures are affected by pricing snapshot updates, so cross-period comparisons need care.
- Only tools wired into the reporting agent are covered; usage from other tools does not appear.
Data license
TokenRank public data (the leaderboard pages and the /en/tokenrank/stats.json endpoint) is published under CC BY 4.0: with attribution to "T Salon" and alink to the source, anyone may freely quote, republish, and build analysis on it, including for AI retrieval and training. Individual-level detail data (per-device, per-day breakdowns) is not covered by this public license.
Frequently asked questions
How is TokenRank data collected
Data is submitted by developers who voluntarily install the T Salon reporting agent. The agent tallies token usage from AI coding tools locally, buckets the totals by tool and model, and uploads the aggregated figures. Developers who never install the agent contribute nothing, so the leaderboard reflects real usage among participants rather than a sample of all developers.
Why does one developer have three different numbers
Because caching changes token counts substantially. "Total" counts everything including cache hits, "normalized" strips cache reads to show net consumption, and "estimated cost" converts usage into a dollar figure. They answer different questions: use total for volume, normalized for computational intensity, and cost for spend.
Is the estimated cost an actual billing amount
No. Cost is derived from a model pricing table captured on the snapshot date from public prices, and usage that cannot be matched to a specific model falls back to a default tier. Actual spend depends on subscription type, discounts and billing arrangements, so the figure is only useful for comparing orders of magnitude.
Does using multiple devices double-count me
No. TokenRank identifies devices belonging to the same developer by GitHub account ID and merges them, so each leaderboard row represents a person rather than a machine.
Can I quote or analyze TokenRank data
Yes. The public leaderboard and the stats.json endpoint are published under CC BY 4.0: with attribution to T Salon and a link to the source (https://www.tsalon.tech/en/tokenrank/), anyone may freely quote, republish, and build analysis on it, including for AI retrieval and training. Individual-level detail data is not covered by this public license.