A popular Codex pattern is simple: let Sol plan the work, then delegate execution to much cheaper Luna subagents.
The price gap is real. On the current Codex token-based credit rate card, Luna's input, cached-input, and output rates are roughly one twenty-fifth of Sol's.
That does not make the whole workflow one twenty-fifth of the cost.
Every delegation creates more work around the cheap model. The main agent has to split the task, package context, monitor progress, receive the handoff, check the result, and sometimes retry. Subagents also use their own model calls and tools.
One Reddit user compared similar refactoring work and reported that a Sol plus Luna setup finished roughly 20 to 40 percent faster while costing about 35 percent more overall. That is one person's test, not a universal result. It is still a useful warning: the model rate is only one line on the bill.
The routing question comes first
The wrong question is: "Which model should every subagent use?"
The useful question is: "Should this task be delegated at all?"
I built Token Saver as a routing discipline for that decision. It is not an automatic dispatcher and it does not bypass usage limits.
The process starts with a hard filter: which models are allowed to see the data?
Then it scores five practical dimensions:
- Uncertainty: how much judgment is required?
- Error cost: what happens if the result is wrong?
- Verifiability: can a machine or reviewer check the output cheaply?
- Context complexity: how much shared state must travel with the task?
- Side effects: can the action be reversed?
Clear, repeatable, testable, and reversible execution can start on a lower-cost tier. Ambiguous or irreversible judgment stays on a stronger model. Mixed tasks should be split so execution and judgment are priced separately.
"Always choose the cheapest" failed my own evaluation
In a 30-task evaluation published with the project, dynamic routing used an estimated 0.30x of the strong-model baseline and completed 89.0% successfully.
A fixed mid-tier policy used 0.34x and reached 96.7% on the same task mix.
| Policy | Estimated relative cost | Success rate |
|---|---|---|
| Dynamic routing | 0.30x | 89.0% |
| Fixed mid tier | 0.34x | 96.7% |
This is a small project evaluation, not a promise about every workload. The interesting result is that the more elaborate routing policy was slightly cheaper but less successful. No fixed answer wins on cost, reliability, and complexity at the same time.
Count the workflow, not the cheapest worker
A useful cost review includes:
- main-agent planning and monitoring
- context passed to each subagent
- retries and verification
- tool calls
- the human time needed to inspect failures
The goal is not to minimize every token. It is to reserve expensive models for the judgment that only they can do.
Sources: OpenAI subagent documentation, Codex rate card, and the individual Reddit cost report.
AI-assisted drafting. Project metrics, rate-card claims, and anecdotal boundaries were reviewed before publication.
Top comments (0)