Documentation · AI & cost

Cost & analytics

Every AI call the platform makes is metered — model, agent, token counts, and exact cost — and attributed to the ticket it served. Nothing about your spend is a mystery.

Where to look

PlaceShows
Project → AnalyticsTotal spend, daily spend chart, spend by agent and by model, board status distribution, and the cost-per-ticket ranking.
Any Done ticketIts full AI cost: every generation, repair, and enrichment call it took.
Project → Settings → Model UsageAll-time project totals, broken down by agent and by model.
Admin → Recent invocationsThe raw ledger, call by call — with separate token columns for uncached input, cache writes, cache reads, and output.

Reading the token columns

Providers bill input tokens in three separate buckets, and the ledger shows all of them:

  • In — uncached prompt tokens at the full input rate.
  • Cache write — prompt stored for reuse (small one-time premium).
  • Cache read — a repeated prompt served at ~10% of the input rate. Repair loops live here; it's where the platform saves you money.
  • Out — generated tokens, the most expensive bucket. This is the real cost of code generation; the platform minimizes it by having repairs return surgical edits instead of whole files, and document refinements return only changed sections.
Rule of thumb from real usage: a typical feature ticket lands for $0.30–0.50 of AI spend end-to-end (generation + validation repairs), and a merge-triggered documentation update for about $0.10 — half that with the batch discount on.

Controls

  • Monthly budget cap (Admin) — hard stop for BYOK spend.
  • Prepaid balance — platform-model usage draws down transparently; top up from Admin.
  • Batch tier toggle (Admin) — 50% discount on background document work.
  • Free tiers — route agents to self-hosted models or your Claude subscription for $0 development.
Cost & analytics — CogniForge Docs