Cut Codex Costs: Assign Strong Model for Planning, Cheap Models for Execution & Review

Codex is a powerful AI coding agent, but its per-task cost can be high because a single request actually triggers a chain of model calls. The article breaks down this chain into four distinct roles:

Main model : understands the goal, plans the task, coordinates execution, and decides the next step. Errors here propagate downstream, so this role needs strong reasoning.

Sub-agent : performs concrete actions—editing files, writing code snippets, running tests. Tasks are already well-defined, so top-tier reasoning is not required.

Review model : checks code for issues and suggests fixes. The scope is narrow, making it a good fit for a cheaper model.

Compact : compresses context when it grows too long. This step currently cannot be assigned a separate model; it follows the main model.

The core insight: use a high-capability model only for planning, and offload execution and review to lower-cost models . Once configured, Codex applies this division of labor automatically for every task.

Step 1: Configure Main Model and Review Model

Edit the global configuration file:

macOS/Linux: ~/.codex/config.toml Windows: C:\Users\<username>\.codex\config.toml Add the following entries (example uses OpenAI's gpt-5.6 series):

model = "gpt-5.6-sol"
model_reasoning_effort = "high"
review_model = "gpt-5.6-luna"

model

is the main planner; model_reasoning_effort = "high" enables deeper reasoning. review_model handles code review with a lighter model.

Step 2: Configure Sub-agent Model

Create a worker configuration file at ~/.codex/agents/worker.toml (create the directory if needed) with:

model = "gpt-5.6-luna"
model_reasoning_effort = "medium"

The sub-agent inherits the session's provider, so the model must be supported by the same provider. Using gpt-5.6-luna with medium reasoning effort cuts cost for the high-volume coding and test-running steps.

Using a Third-Party Provider (Volcengine Agent Plan)

If you prefer a non-OpenAI provider that supports the Responses API, you can define a custom provider in config.toml:

model = "<strong-model-in-plan>"
review_model = "<lightweight-model-in-plan>"
model_provider = "volcengine-agent-plan"
model_supports_reasoning_summaries = true

[model_providers.volcengine-agent-plan]
name = "volcengine-agent-plan"
base_url = "https://ark.cn-beijing.volces.com/api/plan/v3"
env_key = "ARK_API_KEY"
wire_api = "responses"

Then in worker.toml set the sub-agent to the plan's lightweight model:

model = "<lightweight-model-in-plan>"
model_reasoning_effort = "medium"

Note: Agent Plan requires a dedicated API key (not a standard Volcengine Ark key). Model names must match those available in your purchased plan.

Limitation: Compact Step

The context-compression step ( Compact) cannot be assigned a separate model; it always uses the main model. Since the bulk of token consumption comes from execution and review, optimizing those two stages still yields significant savings.