Operations

What Kyle costs to run

Every pipeline run emits a structured event at each agent boundary, with token counts and a cost computed at emit time from a versioned pricing table. This page is that data, aggregated over a rolling window. Nothing here is estimated.

3 of 3 runs are recorded demos

Genuine pipeline runs with genuine cost and latency, recorded to seed this page rather than started by a visitor. They are counted in every figure below. As real traffic arrives this share falls.

Runs (30d)

Count of pipeline runs that reached a terminal state.

Success rate

100.0%

Runs delivering a verified document ÷ all runs. Verifier rejections count as failures.

Research coverage

100%

Documents backed by at least one web search ÷ documents produced. The remainder either declared the topic self-contained or were written from model priors with no stated reason.

Degraded

Runs that delivered a document but fell short of the research contract — unresearched, uncited, or citing a source that was never fetched. Counted apart from both success and failure.

Cost per document

$0.0099

Mean over every run that produced a document, degraded ones included — they cost real tokens. Token counts × the pricing table below, computed at emit time.

p95 cost per document

$0.0107

Nearest-rank p95, so the figure is a cost some run actually incurred.

p50 latency

21.4s

End to end: request accepted through document uploaded.

p95 latency

30.2s

Nearest-rank over all terminal runs. Tails, not averages.

p99 latency

30.2s

Nearest-rank. With fewer than 100 runs this is the slowest observed run.

Total spend (30d)

$0.03

All runs, including failures — failed work still burns tokens.

Daily cost

One point per day across the window. Days with no runs are zero, not gaps.

2026-08-082026-09-06

Cost and tokens by agent

Attribution, not just a total. Each span records its own token counts and price.

Agent	Calls	Tokens in	Tokens out	Cost	Share
planner	6	11.5K	563	$0.0049	16.3%
researcher	3	0	0	$0.0000	0.0%
writer	3	13.5K	7.3K	$0.0222	74.4%
verifier	3	8.8K	60	$0.0028	9.4%

Failure taxonomy

A closed enum, so every class is listed whether or not it fired. Runs ending in a terminal failure do not consume the user’s daily quota.

internal0
invalid_citation0
malformed_json0
no_research0
provider_http0
provider_keys_exhausted0
provider_rate_limit0
provider_timeout0
render_failure0
scraper_failure0
uncited_document0
upload_failure0
verifier_rejected0

Pricing table

Version 2026-08-09.2. Cost is computed when the call happens, from these rates — not recomputed later at whatever the provider charges today.

Model	Input / 1M	Output / 1M	Effective
gemini/gemini-3.5-flash-lite	$0.30	$2.50	2026-08-09
gemini/gemini-3.5-flash	$1.50	$9.00	2026-08-09
gemini/gemini-3.6-flash	$1.50	$7.50	2026-08-09
gemini/gemini-2.5-flash	$0.30	$2.50	2026-08-09
kimi/kimi-k2.6	$0.95	$4.00	2026-08-09

Generated 2026-09-06 19:41 UTC, revalidated every 15 minutes. Cost instrumentation landed 2026-08-09; figures cover runs recorded since then, so a full 30-day window is only complete from 2026-09-08. Aggregates only — no per-user data is recorded or shown.