Thinking and Reasoning Effort
Use the provider's real reasoning channel instead of asking the model nicely in a prompt — and pay for exactly as much of it as you want.
Most current models have a real reasoning channel: a separate space the model thinks in, with a budget you set and tokens you are billed for. SynapseOrch exposes that as one setting.
| Effort | Use for |
|---|---|
off | Routing, classification, extraction — work with no thinking to do |
low | Ordinary tool-using turns |
medium | The default |
high | Planning, multi-constraint reasoning, hard debugging |
Set it per agent, with Reasoning in the agent editor. An agent that sets nothing — Default in the editor, which names the level — inherits the workspace's Reasoning Effort, in Settings → Models (medium unless changed). A Claude CLI model with no setting anywhere runs at low. There is no per-step setting: a workflow step uses its agent's effort.
Why this replaced a prompt convention
Before this, "reasoning" meant the system prompt asked the model to wrap its thinking in [REASONING]...[/REASONING] markers and the engine stripped them back out afterwards.
That is a convention, not a capability. It works exactly as well as the model's willingness to follow it — which is why it read well on Claude and badly on Gemini. Models that have a real channel now use it; models that do not keep the prompt convention, so nothing regresses.
What Off does
Not every model can stop thinking, and Off does the least each one allows. The agent editor and Settings → Models say what Off does for the chosen model.
| Model | Off |
|---|---|
| Gemini 2.5 Flash | Stops thinking (a budget of 0) |
| OpenAI gpt-5.x | Stops thinking (none) |
| OpenAI gpt-5, gpt-5-mini, gpt-5-nano | minimal, their lowest level — measured to think nothing |
| Gemini 3 Flash | MINIMAL, its lowest level, which may think a little |
| Gemini 2.5 Pro | Its 128-token floor |
| Gemini 3 Pro, Claude 5, OpenAI's o-series | No Off: the least they think is Low. The editor offers no Off for them, and an Off already set runs as Low |
On Gemini, Off cuts output tokens by about thirty times against the default. The setting is a preference: a model that cannot honour it still runs, at the least it allows.
The response cap
Max Response Length caps what a model may write in one response, thinking included, in tokens. Set it for the workspace in Settings → Models, and per agent in the agent editor, where it wins. Blank is automatic: 8,192 for most providers, and no limit for OpenAI, OpenAI-compatible and local servers.
A model that thinks always gets room past its thinking: Anthropic and Bedrock reject a thinking budget at or above the cap, so SynapseOrch raises the cap along with the budget rather than letting the call fail. Gemini 3's cap is sized to its thinking level, and Gemini's to at most 65,536.
Cost
Reasoning tokens are billed, and high can cost several times off on the same prompt. They appear in Usage like any other tokens.
The cheapest workflow is usually not the one that thinks least everywhere, but the one that thinks hard at the step that decides something and not at all in the steps that merely move data. Give those steps different agents — one with high, one with off. LLM, Tool and Evaluator steps have no agent of their own and think at the workspace's setting (medium unless changed) on models that have a reasoning channel.
Where the model has no channel
Effort is ignored where the model has no native reasoning channel; those models keep the [REASONING] prompt convention. The agent editor shows the Reasoning control only for models with a channel, so you are not guessing whether a setting is doing anything.