Network statistics

The public cloud samples every routed inference response — time to first token, throughput, token counts, GPU attribution — and publishes anonymized aggregates on permissionless endpoints. No API key, no account data, no prompts: only counters and timings, grouped by GPU model and served model.

Snapshot — September 18, 2026

In the analytics pipeline's first day, the mesh served 48.4 million input tokens and 374 thousand output tokens across 563 requests, with 0 cached input tokens reported so far.

Performance over the same window: 100% success rate, ~102 output tokens per second on average (p50 103), time-to-first-token of 623 ms at p50 and 809 ms at p90.

A note on the cached column: prompt-cache accounting is wired end to end, but an upstream only reports cache hits when it exposes them — none of the currently routed providers do, so it reads 0 until one does.

Usage totals apply the same settlement predicate as billing (success_only=1 by default): only responses whose usage the credit ledger would settle are counted, so the totals line up with what buyers are actually charged.

Fetch live numbers

The snapshot above is static; the endpoints always serve current aggregates:

# performance percentiles + daily token usage, last 30 days
curl "https://api.opentela.ai/v1/leaderboard?hours=720"

# today only
curl "https://api.opentela.ai/v1/leaderboard?hours=24"

# include aborted/failed responses in the usage totals
curl "https://api.opentela.ai/v1/leaderboard?hours=720&success_only=0"

The response pairs entries (GPU × model performance percentiles) with usage (per-day token totals). window_days is the requested window rounded up to whole days, capped at 30 — the raw-sample retention.