RESEARCH · PLAN

When Agents Report
Their Own Cost

A planned study of the gap between self-estimates, rate-card totals, and invoices.

Origin
Gaia Research · Open study
Skill
cost
Status
PRP PROPOSED

SIMULATED LOCAL RUN

Trace the tokens. Price every turn.

EXAMPLE DATA

A documented `cost` run reads a local JSONL session, includes cache reads and writes, and recomputes USD from the pinned LiteLLM rate card. Nothing is executed or uploaded from this page.

$ /cost --latest --offline

-- pi  019f4e66-4216-74cc-834b-6633b275a278
   cwd:    C:\Users\me\projects\web
   time:   2026-07-10T23:39:12.278Z  ->  2026-07-10T23:57:10.128Z
   tokens: in=3,552 out=19,986 cache_r=843,710 cache_w=72,737 total=939,985
   cost:   $1.6104

------------------------------------------------------
1 session(s)   tokens=939,985   cost=$1.6104

Example output format and values are drawn from the public `skill-cost` documentation.

Gaia Research · Planned study · No measurements reported yet


Abstract. Agents are routinely asked what a session cost, yet their answer can omit cache-read tokens, rely on a stale rate card, or lose pre-compaction turns from view. This planned study will compare each agent self-estimate with a complete rate-card total and the matching vendor invoice. It is a research plan, not a completed result: no measurements, distributions, or conclusions are reported here.

1. Research question

Agents systematically under-report their own token cost. How much, and why?

The study will measure the signed delta between an agent’s stated USD estimate and invoice truth, then determine whether the direction and size of the gap are systematic across sampled orchestration sessions.

2. Planned comparison

ComparisonPlanned sourceRecorded value
Agent self-estimateSession response after the cost questionReported USD and stated assumptions
Rate-card totalComplete JSONL trace priced by skill-costPer-session USD, model mix, cache reads, and pricing snapshot
Invoice truthVendor invoice for the matching billing windowActual billed USD and reconciliation notes

3. Planned failure-mode attribution

For every under-reported session, the analysis will separate the gap into:

  1. cache-read tokens omitted from the estimate;
  2. price changes after the agent’s training cutoff or a stale local rate card; and
  3. pre-compaction turns invisible to the post-compaction agent.

The pricing catalog will be pinned for every run. LiteLLM’s model_prices_and_context_window.json is the intended canonical catalog; its exact revision will be recorded with the dataset.

4. Planned artifacts

ArtifactLocationStatus
Sampling and reconciliation protocolMETHODOLOGY.mdPlaceholder
Redacted session snapshots and comparison rowsdata/Not collected
Estimate-versus-truth chartscharts/Not generated
Deterministic analysis and chart scriptsscripts/Not written

Privacy: the eventual dataset will contain only redacted, consented snapshots. It will not include session contents, credentials, or invoice account details.