Token use is concentrated among Cursor’s most engaged users. Over the last four weeks, the top 10% of users accounted for nearly two-thirds of all tokens, while the remaining 90% accounted for just over one-third.
We compared these cohorts across language, model choice, parallel agent use, and feature adoption to see how their patterns of work differ.
The top 10% of users consume nearly two-thirds of all tokensShare of users and token usage over the last four weeks
- Top 1%
- p90–p99
- p50–p90
- Bottom 50%
Language mix varies by cohort
The languages represented in accepted agent diffs vary by engagement cohort. Over the last 28 days, the most engaged users had a larger share of accepted diffs in Rust, Go, and TypeScript. The distribution was more even among users in the p50–p90 band.
Language is an imperfect measure of the underlying work. A TypeScript file, for example, could contain a React component or server code. The chart describes differences in language mix without establishing what kinds of tasks produced them.
The most engaged users work in a different mix of languagesOver/under index (1.0 = global share of accepted diffs)
- p50–p90
- Top 1%
Token use varies by language
To see if some of the difference in token volume could stem from different languages being more or less token efficient, we looked at the token volume per conversation broken down by language. Systems and backend languages skew heavier with a few exceptions.
Median token use varies widely by languageMedian tokens per conversation
Some of this variation tracks the number of requests in a conversation, though the relationship is not uniform. C/C++ had high median token use and a high number of requests, while Rust also had high token use, but across fewer requests.
Languages with similar token use can have different request patternsMedian tokens vs median requests per conversation
- Compiled
- Interpreted / scripting
- Markup / templates
Model choice varies by language
Model use also varies by language. We compared each model’s share within a language with its overall share across all languages.
Within the Anthropic model family, Opus was overrepresented in Rust, C++, and Go. Fable remained closer to its overall share across most languages. These results show an association between language and model choice, though they do not indicate whether either one drives the other.
Model choice varies by languageOver/under index (1.0 = model's overall share)
- Opus
- Sonnet
- Fable
Power users run agents in parallel
The clearest difference between engagement cohorts appears in how often users run concurrent agents.
Among the most engaged users, 86.6% opened a new agent chat while at least one other agent was running, and 40.4% did so while at least two others were running. Those figures were 25.2% and 3.1% for users in the p50–p90 band.
The most engaged users are much more likely to run agents in parallel% of users (last 28 days)
- ≥1 other running
- ≥2 others running
Even within the same engagement cohort, parallelism varies by language. Among users in the p90–p99 band, those working in Go and TypeScript were more likely to run agents in parallel than those working in HTML or CSS.
Parallel agent use is more common in some languages than others% ever opened a chat while ≥1 other agent was running
Power users work differently
These patterns point to a different way of working with agents. The most engaged users choose different models, run more work in parallel, and make greater use of features built for multi-step tasks. The language split may reflect where this way of working is easiest today, in environments where an agent can write code, run tests, and respond to the result with less human input.
As agents become capable of completing and checking more work on their own, we’ll track whether this way of working spreads to more users and kinds of tasks. Today’s most engaged users may offer an early view of how agent use develops across the rest of the user base.