Who pays for agent work
Economic itself is free and runs on your Mac. Agent work runs on the official Codex CLI signed in to your own ChatGPT account, so it counts toward that account’s Codex usage limits and plan. Economic does not bill you, resell model access or hold an API key.
Economic’s internal credits are an accounting tool for agents’ budgets. They have no monetary value and are never charged to you.
What a task uses
These figures come from 99 completed worker tasks in Economic’s own benchmark (games, research, papers, analysis, code repair, documents and simulations), all run with the default model gpt-6.1-sol at xhigh reasoning effort.
| Measure | Typical (median) | Heavy (90th percentile) | Largest seen |
|---|---|---|---|
| Input tokens | 550,000 | 2.2 million | 6.1 million |
| Of which cached | 93% | 97% | 97% |
| Output tokens | 18,000 | 25,000 | 43,000 |
| Time | 8 minutes | 13 minutes | 45 minutes |
Most input is cached context that agents re-read between steps. How much of your plan a task uses depends on how OpenAI meters Codex for your account, which can change; check your usage in ChatGPT’s Codex settings.
What multiplies usage
- Concurrent workers. Capacity is automatic by default. Two workers at once use roughly twice as much in the same time. Set a fixed limit in Economic’s settings or with
economic config set harness_concurrency 1. - Delegation. A coordinator can fund helper agents for larger goals; each helper is its own model session.
- Retries and follow-ups. A failed check or a follow-up message starts more model work. Economic never replays an attempt whose outcome is unknown.
- Long browsing sessions. Screenshots and page text add input tokens.
Keep usage predictable
- Start with one outcome and one worker, and check the result before queuing more.
- Use the off switch in Settings to stop new work; it does not start or cancel anything by itself.
- Choose a lighter model or effort for routine work in Settings → Defaults when available for your account.
- Session credit limits cap how much a single conversation can allocate to agents.
What does not use your plan
Installing Economic, setting it up, opening the app or plugin, reading status, browsing saved results and drafting messages start no model work.