Repository navigation
Add a Smart model picker option that automatically routes each task to the right Codex model #43868
Description
What variant of Codex are you using?
Codex App, CLI, and ChatGPT Work
What feature would you like to see?
The problem
Choosing a model is an unnecessary decision point on every task. Frequent Codex users have to guess whether a task needs the strongest model or whether a lighter model would be enough. The result is either spending premium usage on simple work or underpowering a complex task and needing retries.
Proposed feature
Add Smart as an option in the model picker, plus an optional /smart prompt command for one-off use.
Before a task is dispatched, Smart would select the best available model and reasoning level from task signals such as task type, repository scope, ambiguity, expected tool use, and estimated complexity. It should be available consistently in the Codex App, CLI, and ChatGPT Work.
Important behavior
- The routing decision should not consume a separate user turn or meaningful extra usage. A lightweight local/rule-based classifier or platform-side router could make the choice.
- Respect user controls: a maximum-cost/maximum-capability setting, a default preference, and an easy per-task override.
- Show a small, non-intrusive explanation after routing (for example, “Smart chose GPT-5.6 Terra — focused code change in a small repo”) so the choice is understandable.
- Escalate only when justified, rather than silently using the most expensive model by default.
- If Smart is uncertain, choose a sensible default and let the user change it before or after the task.
Why this matters
This would make Codex easier to trust for everyday use: users can focus on their actual work, conserve their limited premium usage for work that genuinely needs it, and still get stronger reasoning for difficult multi-file, debugging, or agentic tasks. It is especially valuable for students, individual developers, and teams that use Codex frequently but do not want to become experts in model tradeoffs.
Additional information
I started exploring this as a standalone “Smart Model Selector,” but the product itself is the right place for it because only Codex has the full context on model availability, capacity, and usage economics. The ideal experience is a durable Smart picker setting for all tasks, with /smart as a convenient one-off override.
Activity
MrCsabaToth commentedon Sep 18, 2026
Very much needed!!!
kehator commentedon Sep 23, 2026
I’d like to support this request and suggest an explicit, lightweight way to invoke Smart routing: let users start a prompt with a model name, such as “Luna — find information about…”, “Sol — analyze this file…”, or “Astra — investigate this serious bug…”. Codex could recognize the prefix in typed or voice-transcribed input, select that model for the task, and show the selection before proceeding. This would make model switching semi-automatic while keeping the choice and usage tradeoff in the user’s hands.
vertex-tech-labs commentedon Oct 12, 2026
Proposal extension: user-defined optimisation priorities for Smart model routing
I support this request and would like to propose a Smart Balanced Mode that lets users control what the automatic router optimises for, not just whether automatic selection is enabled.
My workflow spans software architecture, implementation, debugging, product design and research. Those activities differ greatly in reasoning requirements, and manually selecting models disrupts workflow and can exhaust usage allowances unnecessarily.
Three user-selectable priorities:
Additional capabilities I would value:
For example, I may plan a complex system, implement a straightforward change, run tests and then investigate a difficult defect in the same work session. A single fixed model/effort setting is unlikely to be optimal for all four stages.
This is about making better use of the allowance already included, not simply increasing usage limits. A configurable automatic router would especially help independent developers and small teams.
Related ideas: #46366 (cost-aware routing) and #47576 (dynamic model and effort routing).