Token usage

Scout’s model calls run on your own Codex or Claude Code login, so they count against that subscription’s usage allowance. This page gives measured token counts and times, not prices. Every number below says where it comes from and when it was measured. Yours will differ: run gigai scout metrics for your own.

What an agent costs

Two different things use tokens when you drive Scout from an agent:

• Scout’s own model calls: ranking, assessing, tagging titles (and tailoring, which is switched off in 0.1.11). This page is about these. Scout records each one on your computer (kind, model, tokens, seconds; no text).
• Your agent’s own conversation: the grid it reads, your questions and its replies. That is your agent’s usage. GigAI does not see or measure it.

What keeps Scout’s side under your control:

• It asks before it assesses. gigai scout new (and Assess these in the UI) gives the count and an estimate from your own past calls, and assesses only on a yes. One model call per posting, at most 50 postings per approval, and a running batch can be cancelled (Assess a batch).
• gigai scout metrics shows the averages once the first calls are recorded: tokens, seconds and error rate per kind of call and per model. A token estimate is shown only after a call has been measured.
• Daily caps for background work. The background pipeline is off by default in 0.1.11. A home that switched it on explicitly makes at most 40 model calls a day with it. Background ranking makes at most 100 calls a day, with a warning past 60. Both are in Settings > Background pipeline. A call that would go over the cap is not made.
• The first run is the expensive one. The first gigai scout new --yes catches up on every new posting at once. After that, a day is a few dozen new postings at most.

Tokens per step

One call each, on the same made-up resume and two short made-up postings, through GigAI’s own adapters on a scratch install. “Tokens” is input plus output; “cached” is the part of the input the provider served from its prompt cache.

Step	What one call covers	Codex CLI	Claude Code CLI
Rank	50 postings	11,092 tokens (9,446 in, 0 cached; 1,646 out), 45 s	7,100 tokens (4,709 in, 0 cached; 2,391 out), 13 s
First assessment	1 posting (average of 2 calls)	11,630 tokens (10,948 in; 682 out), 21 s	11,977 tokens (10,465 in, 531 cached; 1,512 out), 11 s
Tag titles	50 titles	8,836 tokens (8,125 in, 3,840 cached; 711 out), 26 s	3,586 tokens (2,547 in, 0 cached; 1,039 out), 6 s
Tailor	1 resume for 1 posting	10,338 tokens (9,974 in, 0 cached; 364 out), 14 s	9,652 tokens (8,929 in, 531 cached; 723 out), 6 s
Re-assess after tailoring	1 posting, the tailored resume	11,838 tokens (11,016 in, 3,840 cached; 822 out), 24 s	12,197 tokens (10,633 in, 531 cached; 1,564 out), 11 s
Scout ATS score	1 resume	0 tokens: computed on your computer, no model	0 tokens: computed on your computer, no model

The two tailoring rows were measured on 0.1.10. Tailoring and the background pipeline are off by default in 0.1.11, so a default home makes neither call.

Source: GigAI 0.1.10.8 development build, 2026-10-03, codex-cli 0.159.3 and Claude Code 2.1.288, each CLI’s default model, six calls per CLI, read with gigai scout metrics --json. Of the two Codex first assessments, one had no cached input and one had 3,840 cached tokens. The Claude Code CLI ran the rank and tag calls on claude-sonnet-5-5.

How to read it:

• These postings were short. A real posting is several times longer, so a real assessment uses more input tokens than this table shows. See the next section.
• Most of an assessment’s input is not your text. Each CLI adds its own instructions to every call, and Scout’s prompt carries the rules and the answer format. That part repeats on every call, which is why a provider can cache it.
• A job you answered a question for costs one more call: the re-assessment, when you ask for it. The pick of the resume, the Scout ATS score and the Scout label are local. (A home that switched the background pipeline on explicitly also pays for one tailoring call.)
• One rank call covers 50 postings, so ranking is cheap per posting, and a posting is ranked once for a given resume and settings.
• Six calls are a sample, not an average. Time per call moves with the provider’s load.

The Claude Code CLI also reports an API price for each call (0.02 to 0.06 US dollars per call in this sample). On a subscription that is not what you are billed; it counts against your allowance. The Codex CLI reports no price.

A real run: the first catch-up and a normal day

Real numbers from the maintainer’s own install: GigAI 0.1.10.7, Codex CLI on a ChatGPT subscription, default model, real postings, read on 2026-10-03 with gigai scout metrics --json.

Kind	Calls	Average per call	Cached share	Time per call
Assess	322 (1 failed)	23,526 tokens (22,027 in; 1,499 out)	13,014 of the 22,027 input tokens (59%)	42 s
Rank	30 calls for 1,433 postings	21,079 tokens (19,431 in; 1,648 out)	12,105 of the 19,431 input tokens (62%)	45 s

What that meant for the subscription, as that user read it on the same day (examples, not a promise for your plan):

• About 180 assessments and 4 rank calls used about 3% of the Codex allowance: roughly 0.017% per assessment.
• The first catch-up, 333 postings, used about 5 to 6% and took about an hour (42 seconds each, 4 at a time).
• A normal day is a few dozen new postings. Worked out from the numbers above, 30 assessments are about 0.7 million tokens, about 5 minutes, and about half a percent of that allowance.

Two things changed in 0.1.10.8 that make these 0.1.10.7 numbers an upper guide for Codex: Scout now turns every Codex tool off, so Codex no longer sends its tool definitions with each call. On a made-up prompt during development (2026-10-03) the input of one Codex call fell from about 33,800 tokens to about 7,000. How much a real assessment falls has not been measured yet.

Learning pathways

Measured on the 0.1.11.10 head, from the cost record of each run:

What	Calls	Tokens	Time
A course for a role (real: MLOps 48 calls, AI forward deployed engineer 43 calls)	about 43 to 48	about 0.86 to 1.0 million (MLOps about 1.02M, AI forward deployed engineer about 0.86M)	about 2 to 2.5 hours of background work (130 to 140 minutes)
The pathway page of a shipped course	about 1	about 32k	about 90 seconds

A course uses about 1 million tokens. On an entry-level plan, such as Claude Pro or ChatGPT Plus, that can use up most or all of your weekly limit; use a higher plan, or one of the courses that come with GigAI. See Learning pathways.

A course also makes about 1,300 page fetches to public documentation sites. Scout asks first and stops a course by itself after 120 calls or 1,500 fetches. See Learning pathways.

Codex or Claude Code

From the six-call sample above, for the same input:

• Tokens per assessment were about the same (11,630 on Codex, 11,977 on Claude Code).
• Claude Code answered faster (11 s against 21 s for an assessment, 13 s against 45 s for a rank call).
• Rank and tag calls used fewer tokens on Claude Code (7,100 against 11,092, and 3,586 against 8,836): Scout calls Claude Code in a lean mode for those, with a one-line system prompt.
• What a token is worth differs per plan. A share of an allowance is only known for the Codex example above. No Claude Code allowance share has been measured.

See your own numbers

gigai scout metrics                  # averages per kind of call and model
gigai scout metrics --kind assess    # one kind: assess, rank, tag, tailor, extract or interview
gigai scout metrics --json           # the same, for a script or your agent

In the UI the same table is under Settings > Model usage. Everything is read from a file on your computer; nothing is sent anywhere to show it.

gigai scout pipeline status shows today’s background calls against the daily cap.