Allow users to assign a capped compute ticket to a Codex task, with live tracking and automatic stop

What variant of Codex are you using?

Codex in the ChatGPT desktop app on macOS

What feature would you like to see?

Summary

Add an optional per-task or per-conversation budget to Codex. A user could assign a maximum amount of credits to a piece of work—and, when the conversion is reliable, display its value in euros or the user’s local currency—then monitor usage and automatically stop additional paid consumption at the chosen cap.

The intended interaction is simple:

Give this task a boost worth no more than €20.

Codex would then show the scope, credit budget, available estimate, settings that may accelerate consumption, and the stop rule. Work would begin only after human confirmation.

Problem observed

I am an advanced individual user without a strong computer-science background. The base offer already makes complex work possible and naturally creates interest in purchasing additional capacity occasionally to finish an important task.

We tested approximately €100 of additional credits to get past an interruption caused by the five-hour limit and to explore the potential of faster settings. The additional capability was real. However, the credits were exhausted within a few hours without sufficient visibility into:

• which conversation or task was responsible;
• the consumption rate;
• the effect of the selected settings;
• how much work could still reasonably be completed;
• when usage would move from the allowance included in the subscription to purchased credits.

This lack of control discourages further purchases. A user may be willing to spend €10, €20, or €50 on an important task, yet refuse to purchase again when consumption cannot be predicted, monitored, or stopped.

Proposed behavior

1. Define the ticket — The user chooses a cap in credits or selects a small amount displayed in their currency when the service can guarantee the conversion.
2. Explicit scope — The ticket is restricted to one clearly named task, conversation, or goal. It does not become a general budget available to other work.
3. Preview — Before confirmation, Codex shows the intended work, reasoning effort, potentially expensive capabilities, available estimate, and its uncertainties.
4. Single human authorization — Confirmation authorizes only the displayed cap, task, and settings. Any material change requires renewed approval.
5. Live tracking — The interface shows the initial budget, amount consumed, remaining balance, and depletion rate.
6. Alerts — Codex warns the user when actual consumption diverges materially from the estimate or the cap is approaching.
7. Automatic stop — At the limit, Codex starts no new paid operation and offers to continue with included capacity, pause, or request a new authorization.
8. Outcome report — At the end, the user can see what was completed, what consumed the ticket, and what remains unfinished.

Example flow

Task: finalize the Product dossier
Maximum budget: €20 / equivalent in credits
Scope: this conversation only
Estimate: indicative, with an uncertainty range
Alerts: at 50%, 80%, and upon unusually rapid consumption
At the limit: stop starting new paid operations

[Authorize this ticket]   [Modify]   [Continue without boost]

Required safeguards

• never present an estimate as a guarantee when future consumption cannot be calculated precisely;
• perform currency conversion on the service side and use the credit budget as the reference when necessary;
• disclose that an already in-flight request may produce a small technical overrun, then prevent any new expenditure;
• never renew the ticket automatically;
• prevent a task, subagent, or background continuation from using another conversation’s ticket;
• retain a visible record of the authorization and resulting consumption;
• allow the unused balance to be cancelled immediately or returned to the general credit balance.

Why this would be useful

This feature would not only reduce spending. It could make additional capacity more attractive to smaller users by providing:

• occasional spending tied to a real need;
• bounded financial risk;
• a clearer understanding of the value obtained;
• fewer frustrating interruptions;
• access to advanced capabilities for less technical users without requiring them to understand the full compute model.

The problem is not the availability of additional credits. The problem is allowing users to turn those credits into a controlled decision.

Additional information

This request complements two existing proposals:

• #28985 — Account-wide Codex usage attribution by client, session, and timestamp, which is needed to understand which task consumed the available pool;
• #36590 — Add an explicit token-budget picker for CLI goals, which proposes a token budget for CLI goals.

This request differs in scope: the Codex app, purchased or assigned credits, small amounts that are understandable to users, human authorization, economic tracking, and automatic per-task stopping.

A visual appendix and a more detailed dossier are available. No attachment will accompany the initial creation of this issue, so the request remains focused and no private data is exposed. Sharing those materials later would be a separate action requiring a new review and authorization.

— Philippe Trillaud & Codex

No implementation files, tests, or entry points are named. Start by reviewing the related issues #28985 and #36590, then identify where the macOS Codex app handles task scope, credit usage, authorization, and stopping paid work. Done means the proposed per-task ticket, live tracking, alerts, authorization safeguards, automatic stop, and outcome report are implemented and tested.