What Running PMDI's Agents Actually Costs

Most AI cost coverage is about token prices falling. That's true and also not the number that matters -- your bill is volume × mix × price, and only price is falling. Below is PredictIntel's own real spend, pulled live from the same metering that gates every LLM call this platform makes (api_spend_log) -- not an industry estimate, not someone else's survey. The honest finding, spelled out below rather than smoothed over: per-task token efficiency actually improved over this period, and the bill still grew, because task volume and task mix grew faster than efficiency did.

—

Total spend, all time

—

LLM calls metered

—

$/task, first week → most recent

—

Tokens/task, first week → most recent

—

The mechanism, in PMDI's own numbers

Loading live data…

Weekly spend

By week, since first metered week

Total dollars spent per week across every LLM-backed job on the platform. Hover a bar for the exact figure.

Where this week's spend actually went

Most recent complete week, by generator

Aggregate $/task obscures which specific jobs are expensive. This is the real breakdown — a handful of deeper-analysis jobs (not the high-frequency classification calls) account for most of the dollar spend even though they're a small share of total call volume.

Source	Calls	Cost	$/call
Loading…

Live Intelligence

Recomputed on every page load from PMDI's own live database — no cached snapshot, no third-party estimate.

Loading…