Reasoning Effort
reasoning_effort sets how much the model behind a deepAgent or subagent node reasons before it answers. Lowering it cuts reasoning tokens, cost and time to first answer. none turns reasoning off on the models that support it, which suits cheap subagents that only fetch and format.
It is optional. Left unset, nothing is sent to the provider and the model reasons as it does by default.
Setting it
Add the param to the node. It takes one of none, minimal, low, medium, high, xhigh or max:
{
"id": "deepAgent-4",
"type": "deep_agent",
"data": {
"params": {
"model": { "value": "deepseek-v4-flash-0731", "isExpression": false, "isAttachedToInputNode": false },
"reasoning_effort": { "value": "none", "isExpression": false, "isAttachedToInputNode": false }
}
}
}
The same param exists on subagent nodes, so a main agent and each of its subagents can run at different efforts.
How it is applied
Providers do not accept the same values: one model takes all seven, another rejects none with a 400. The platform records which values each model accepts and checks the param against that list when the workflow is compiled, before any model call.
| Where the model sits | Accepts the value | Does not accept it |
|---|---|---|
model (primary) | Sent to the provider | The run fails at compile time, and the error lists the values the model accepts |
fallback_models | Sent to the provider | The model is called without reasoning_effort, and the worker logs a warning |
So a chain such as qwen3.5-397b-a17b → mistral-large-4-0 → deepseek-v4-flash-0731 set to none works: if the call falls back to mistral-large-4-0, which rejects none, that model is called without the param.
When a long run compacts its history, the summary is written with the primary model's client, so it runs at the same effort.
A value listed below is one the provider accepts: each model was sent all seven values, and the ones answered with 200 are recorded. Accepting a value does not mean the model changes its behaviour for it. For example, llama-3.3-70b-instruct accepts all seven but does not reason. On deepseek-v4-flash-0731 and the Qwen models, none removes the reasoning output.
Accepted values per model
Scaleway
| Model | Accepted values |
|---|---|
deepseek-v4-flash-0731 | none, minimal, low, medium, high, xhigh, max |
deepseek-v4.1-flash | none, low, high, xhigh, max |
glm-5.2 | none, minimal, low, medium, high, xhigh, max |
glm-5.3 | none, minimal, low, medium, high, xhigh, max |
gpt-oss-120b | low, medium, high |
llama-3.3-70b-instruct | none, minimal, low, medium, high, xhigh, max |
mistral-medium-3.5-128b | none, high |
mistral-small-3.2-24b-instruct-2506 | none, low, medium, high |
qwen3-235b-a22b-instruct-2507 | none, minimal, low, medium, high, xhigh, max |
qwen3.5-397b-a17b | none, minimal, low, medium, high, xhigh, max |
qwen3.6-35b-a3b | none, minimal, low, medium, high, xhigh, max |
qwen3.8-27b | none, minimal, low, medium, high, xhigh, max |
scaleway-gemma-4-26b-a4b-it | none, minimal, low, medium, high, xhigh, max |
Mistral
| Model | Accepted values |
|---|---|
devstral-2512 | none, high |
labs-devstral-small-2512 | none, high |
mistral-large-4-0 | low, medium, high |
mistral-medium-2508 | none, high |
mistral-medium-3-5 | none, high |
mistral-small-2603 | none, high |
zai-glm-5-3 | low, high, max |
zai-glm-5-3-flash | low, high, max |
codestral-2508 | Not supported |
ministral-14b-2512 | Not supported |
ministral-3b-2512 | Not supported |
ministral-8b-2512 | Not supported |
mistral-large-2512 | Not supported |
pixtral-12b-2409 | Not supported |
OpenAI
| Model | Accepted values |
|---|---|
gpt-5.4 | none, low, medium, high, xhigh |
gpt-5.4-mini | none, low, medium, high, xhigh |
gpt-5.4-nano | none, low, medium, high, xhigh |
o3 | low, medium, high, xhigh |
gpt-4.1 | Not supported |
gpt-4.1-mini | Not supported |
gpt-4.1-nano | Not supported |
gpt-4o | Not supported |
gpt-4o-mini | Not supported |
o1 | Not supported |
o3-mini | Not supported |
o4-mini | Not supported |
Anthropic and Google
Not supported. These providers control reasoning with their own parameters (thinking, thinking_budget), which the platform does not expose yet. Leave reasoning_effort unset when the primary model is one of them, or the run fails at compile time.
Models: claude-haiku-4-5-20251001, claude-fable-5, claude-opus-4-8, claude-opus-4-7, claude-opus-4-6, claude-opus-4-5, claude-sonnet-5, claude-sonnet-4-6, claude-sonnet-4-5, gemini-3.1-pro-preview, gemini-3.1-flash-lite-preview, gemini-3-flash-preview, gemini-2.5-pro, gemini-2.5-flash, gemini-2.5-flash-lite, gemma-4-31b-it, gemma-4-26b-a4b-it.