Subagents are one of the best features in Claude Code, and they are also an easy way to spend your usage limit faster than you expect. Each subagent is a separate Claude with its own context window, and by default it runs on the same model and effort level as your main session. A few settings let you send the simple work to a cheaper model at lower effort, and one of them is new as of October 31, 2026.
Why subagents cost more than they look
The subagents page says a subagent "sends its own requests, which count toward the same usage limits as your main conversation." The costs page adds that every subagent, and every agent a workflow starts, sends its own requests on top of the main conversation's.
Two defaults make this expensive. When a subagent has no model of its own, it uses the main conversation's model. When it has no effort level of its own and runs on that same model, it uses the effort level of your session. So if you work on the strongest model at high effort, a subagent that only searches files or reads logs runs with the same settings. That is a lot of reasoning power for a job that does not need it.
Setting 1: a cheaper default model for subagents
The model configuration page describes CLAUDE_CODE_SUBAGENT_MODEL as the default model for subagents, agent team teammates, and workflow agents that are not assigned a model another way. It takes an alias such as haiku or a full model name.
There are two details to know. A model named in a subagent's own definition, or a model Claude passes when it starts the subagent, still wins over this default. The variable also does not change the built-in Explore and Plan subagents by itself. To put one model on every subagent, also set CLAUDE_CODE_SUBAGENT_MODEL_FORCE to 1, which needs Claude Code v2.1.257 or later.
Setting 2: a lower effort level for subagents
This is the new one. The changelog for version 2.1.316, released October 31, 2026, says:
"Added CLAUDE_CODE_SUBAGENT_EFFORT to set the effort level subagents run at when nothing else names one, instead of inheriting the session's."
Effort controls how much the model reasons before it answers. The model configuration page says each level "trades token spend against capability" and that lower effort is faster and cheaper for simple tasks. Thinking tokens are billed as output tokens, which is the expensive kind. Before this variable, the only way to lower subagent effort without lowering your own was to edit each subagent's definition.
The changelog does not list the accepted values. I tested low on Claude Code 2.1.322 with one small task and one subagent. With the default settings, the subagent ran on the model of my main chat (Sonnet) at medium effort. With CLAUDE_CODE_SUBAGENT_MODEL=haiku and CLAUDE_CODE_SUBAGENT_EFFORT=low, the subagent ran on Haiku at low effort, and the main chat stayed on Sonnet. The docs do not give token counts per level, so measure the saving on your own work.
Both settings together
Environment variables can go in the env block of your settings.json, so they apply to every session:
{
"env": {
"CLAUDE_CODE_SUBAGENT_MODEL": "haiku",
"CLAUDE_CODE_SUBAGENT_EFFORT": "low"
}
}
The two settings are defaults, so you can still ask for more in the chat. In the same test I added one sentence to my prompt: use the Sonnet model and high effort for this one subagent. That subagent ran on Sonnet at high effort, while the two settings stayed in place.
One behavior changed in version 2.1.317 and matters here. A subagent on a different model from the main conversation no longer takes the session's effort level. It uses that model's saved effort level or its default.
Workflows and single agents
Workflows can start many agents in one run, so they have their own switch. Version 2.1.296 added CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL, which runs every workflow agent on one model while other subagents keep theirs.
For finer control, set model and effort in a subagent's own file. The costs page recommends this directly: "For simple subagent tasks, specify model: haiku in your subagent configuration." The subagents page also notes that a user or project subagent named Explore overrides the built-in one, so you can define your own with model: haiku to run exploration on a lower-cost model.
What I would set, and what I would leave alone
- Keep the main session on the model and effort level the task needs. That is where the decisions are made.
- Send search, log reading, and doc lookups to a cheaper model at low effort. The output of that work is a short summary, and a smaller model writes it well.
- Keep a strong model and higher effort for subagents that review code or verify a fix. A cheap reviewer that misses a bug costs more than it saves.
- Check the result in
/usage. The costs page says the breakdown shows the subagent share, and since version 2.1.299/usagealso warns about max effort and workflows.
The return on this is easy to reason about. If subagents are a large share of your usage and most of them do simple work, moving that share to a cheaper model at low effort gives you more hours of the strong model for the work that needs it, on the same plan. If subagents are a small share, the setting changes little, and /usage will tell you which case you are in.
Work with me
I help people and teams see where their AI and cloud spend goes, then cut the parts that do not pay for themselves. I spent 6 years at Amazon before going independent. If your team uses Claude Code or Codex and the limits or the bill are a problem, book a call at cloudyeti.io/meet.
Book a call