Skip to content
github-actions[bot] edited this page Nov 21, 2026 · 5 revisions

Every model call (#50) appends its cost to [paths] costs (logs/costs.jsonl) with the command and stream it ran for; clippa costs and the war room's Costs panel add that to the Modal run log and your fixed costs: put plans and subscriptions in [costs] monthly ({"Runway Max" = 95}) and a provider's monthly credit in [costs] credits ({"modal" = 30}, shown as credit, not hidden). clippa costs --backfill once reads back model spend earlier runs wrote into their output files (score, segments, receipts, weekly).

clippa costs [--month 2026-10]     # month's spend: models + Modal + fixed, by stage/show/stream
clippa costs --providers           # the Claude API's and OpenRouter's own count, next to our log
clippa modal                       # GPUs up, runs, spend, archive ETA

The providers' own numbers

Our log prices each call from [llm] prices (OpenRouter reports its own cost). clippa costs --providers and the war room's API usage panel ask the providers instead, so spend the log missed (another tool on the same key, the Console's workbench, a crash before the line was written) shows up too. Each is asked at most every 5 minutes.

  • Claude API: the Admin API's cost report, month to date by model. It needs an Admin API key (sk-ant-admin..., Console > Settings > Admin keys; Console organizations only, not individual accounts): set ANTHROPIC_ADMIN_KEY, or put it in secrets/anthropic_admin_secret.txt ([paths] anthropic_admin_key). The normal ANTHROPIC_API_KEY can't read it. Costs appear about 5 minutes after a call. Without the key the panel shows our log for Claude and says what to set; clippa doctor warns.
  • OpenRouter: the key Clippa uses (OPENROUTER_API_KEY or secrets/openrouter_secret.txt): today, this week, this month, all time, what's left of the key's limit if it has one, and the account balance when the key may read it.

Modal's own billing is the real number for transcription (Transcribe on Modal); the bot DMs each time the month passes a multiple of [modal] alert_every_usd.

What things cost

Note

Rough figures measured on our own runs; --dry-run gives the real estimate where a command offers it.

Job Cost
Transcribe an 8 h stream on Modal ~$0.34 (L40S), ~$0.26 (L4)
Weekly issue draft + three articles ~$0.35 + ~$0.10 each
archive ask ~1c
receipts find / script ~$0.20 / ~$0.02
investigate facts ~$0.20 for 43 h; the whole Jer catalog ~$14 once
Search swarm, per lead ~$0.01-0.03
News check (Haiku) well under a cent
Cast photo label ~$0.0002 a photo; Brave image search ~$5 per 1,000
X post ~$0.015 (a link post $0.20, so Clippa never adds one)

One stream, stage by stage

Median spend per stream over this install's logged runs (27 streams, most 7-9 h, October to November 2026; clippa costs has the live numbers):

Stage Median Upper quartile
transcribe (Modal) $0.45 $0.60
--diarize, on top $0.16 $0.18
score (both passes) $0.18 $0.21
segments (one call over the whole transcript) $0.81 $0.92
meta $0.88 $1.05
thumbs (a pick per video, frames as images) $1.01 $1.36
shorts, x about a cent ~$0.04

So a full run on an evening show is about $3.50. A rerun of score, segments, meta or thumbs pays again (score refuses without --force); each of the four prints what a run would cost with --dry-run, from the prompts it would send (four characters a token, times [llm] prices), calling nothing. clippa status ends with the month's spend by stage.

Clone this wiki locally