observed usage, not stars or survey recall
Original data, visible caveats
Every analysis names its sample, update time, definitions, and limitations. Participants are an opt-in public cohort—not a representative survey—and cost figures are API-equivalent estimates unless stated otherwise.
Turns raw leaderboard totals into practical median, quartile, and top-decile reference points with transparent sample coverage.
Read the analysis
Separates developer breadth from token share so a model used by many people is not confused with one generating most tokens.
Read the analysis
Shows both participating-developer reach and token intensity instead of relying on stars, downloads, or directory listings.
Read the analysis
Reconstructs the token composition behind observed cache share and explains where the metric is informative or incomplete.
Read the analysis
Combines observed token composition with dated public pricing to reveal a transparent blended API-equivalent rate.
Read the analysis
Adds a moving average, peak-day context, and daily coverage to the raw trend rather than presenting a context-free total.
Read the analysis
Aggregates exact MCP tool names into server-level evidence while preserving the limits of turn-attributed token measurements.
Read the analysis
Ranks real non-MCP tool invocations and clearly separates call reliability from tokens attributed to the surrounding model turn.
Read the analysis
Places the two agents side by side using the same 30-day cohort, definitions, and API-equivalent cost method instead of opinion scores.
Read the analysis
Cohort, not censusPublic participants choose to appear, so results describe the observed cohort rather than every developer.
Aggregates, not contentMeasurements describe usage totals. Prompts, code, filenames, arguments, and model outputs are not research dimensions.
Attribution mattersSkill and tool tokens belong to the calling model turn; they are not direct execution costs for a tool or MCP server.