CommandAGI · commandagi.com/for/subsidized-tokens

CommandAGI
Ouvrir l’atelier
Menu
Ouvrir l’atelierSe connecter

Subsidized tokens

Each turn goes to the plan with room left.

A Claude or ChatGPT plan is a flat fee. CommandAGI runs your agents on it, reads each sign-in's usage, and holds every sign-in under the cap you set.

Open the workbenchSee the agent CLIs
Réglages → agent / mcp dans le workbench CommandAGI : les CLI d’agents de cet ordinateur, la connexion de chacune et le forfait de chaque connexion.
Recorded in the workbench: the agent CLIs on this computer and the plan of each sign-in.

Before every turn

A cap per sign-in, checked every turn.

agents.usagePolicy in your settings caps each sign-in's share of its plan's window. The host reads the plan's usage and decides before the prompt goes out.

How a turn chooses a planBefore a turn, the host reads the sign-in's last usage reading, from the CLI's own limit event or its usage command. The usage policy compares the window's use with the cap for that sign-in. If it is under the cap, the turn runs and the decision is recorded. If it is at or over the cap, the member is paused until the window resets and the task can move to another member.The plan's usagefive-hour and weekly windowsagents.usagePolicya cap per sign-in or globAllowedevent: policy, then the promptPauseduntil the window resetsHow a turn chooses a planBefore a turn, the host reads the sign-in's last usage reading, from the CLI's own limit event or its usage command. The usage policy compares the window's use with the cap for that sign-in. If it is under the cap, the turn runs and the decision is recorded. If it is at or over the cap, the member is paused until the window resets and the task can move to another member.The plan's usagefive-hour and weekly windowsagents.usagePolicya cap per sign-in or globAllowedevent: policy, thenthe promptPauseduntil the windowresets
json
"agents.usagePolicy": {
  "claude-*": { "weekly": 0.8 },
  "codex-*": { "weekly": 0.8 }
}

Case study · the backlog thread

$831 of API-priced work, on flat plans.

Claude Code reports each turn's cost at API prices. In the backlog thread, its turns on Claude models through Claude plans came to $831 by that measure (case-studies/backlog-thread/stats.json). Codex reports no price for its plan's turns: 4.97 M output tokens on gpt-6-luna and gpt-6.1-sol, with no cost in the record. What the plans cost is not in the record, so this page does not compare the two.

at API prices: 1,120 priced Claude turns (Opus $481, Sonnet $323, Haiku $27)
$831
usage-policy decisions: 2,153 allowed, 2 refused
2,155
pauses at a cap, and 78 more on a stale reading
100
limit readings from the CLIs' own events
1,915
Claude Code's API-priced cost of plan turns, by model
Les chiffres
Claude Code's API-priced cost of plan turns, by model
ModelAt API prices, $
Claude Opus 5.5481
Claude Sonnet 5.5323
Claude Haiku 5.527.1
ProvenanceFile runs/20261023t132208z-0aea (the backlog thread, on this computer)Output case-studies/backlog-thread/stats.jsonTool scripts/sites/case-studies.mjs backlog-threadRan on This computer (dell-gb10-1), reading the run folderWhen

The decision is on the record.

Before the prompt, the policy event: the window, the use, the cap and why. The sign-in's name is left out.

runs/20261023t132208z-0aea/streams/agent-a1@dell-gb10-1real records · 2026-10-23
13:22:31.800 event   policy   cli=claude  decision=allowed  pattern=claude-*  window=weekly  used=0.27  cap=0.8  reading=fresh
13:22:31.863 command agent:task-loop@dell-gb10-1: prompt  cli=claude  model=claude-opus-5-5
13:22:37.364 event   limit    status=allowed  five_hour=0  seven_day=0.26

Use your plans to the cap you choose.

Set agents.usagePolicy once, at your account level, and every host obeys it.

Open the workbenchSee the agent CLIs