Opening

Loading the trace…

One year of LLM traffic: 6.12 billion requests, April 2025 to April 2026

DeepSeek carried most requests in May 2025, and Qwen passed it in March 2026, while the monthly total barely moved.

Requests

–

Input tokens per call

–

ratio of sums, calls with a count

Output tokens per call

–

ratio of sums, calls with a count

Time to first token

–

streaming calls only, seconds

Technical notes: data source, aggregation and limitations

Source. Chutes LLM traffic trace, 6.12 billion requests, rolled up by trace day and hour (arXiv:2608.13573; code and data at github.com/HarvardMadSys/chutes_workload). Day 0 is 11 April 2025.
Aggregation. Every average is a ratio of sums. Tokens per call use calls with a count; time to first token uses streaming calls only.
Comparisons. "From X to Y" claims use full months only (May 2025 to March 2026). April 2025 and April 2026 are partial.
Gaps. 13 January 2026 has 3 of 24 hours and 14 January has 15. Missing hours are shown as gaps, not interpolated; the file gives no cause.
Families. The trace's family grouping. "Unnamed", "Other families" and "Long tail" are kept as the trace groups them.
Clock. Hours are on the trace clock; its time zone is not in the file.
Units. The file does not state the unit of time to first token; it is read as seconds.
Mix and latency. Across 367 days, the day's time to first token and Qwen's share of requests correlate at only 0.19.

Source: Chutes LLM traffic trace (see notes above). Figures are computed from the exported file in this folder.