Model Providers

The active provider is determined by the model name prefix. You can set a global default model in Settings → Models and override it per-agent or per-step.

The screens, step by step: Set up the workspace, in the Guides.

Provider routing table

Reasoning

Many current models have a native reasoning channel, and SynapseOrch exposes it as an effort setting — per agent, or workspace-wide. It is not a per-step setting: a step picks a model, and how hard that model thinks belongs to the agent. Models without a reasoning channel keep a prompt convention instead, and the agent form shows its Reasoning control only for models that have one. See Thinking and reasoning effort.

Setting the default model

Each provider has its own card under Providers in Settings → Models. Click the card, paste the API key (and any region or base URL fields the card asks for) and click the card's Save. Then pick your default in Default model, under Defaults at the top of the page, and click Save changes below the Defaults and Knowledge fields.

A saved key is never shown again. The card says A key is stored with a fragment of it, and its field becomes Replace it — Paste a new key to replace it; the bin button beside it removes the key, after asking.

Under Defaults are two workspace-wide settings that every agent without its own follows: Reasoning effort (on Cloud, Reasoning; see Thinking and reasoning effort) and Max response length, the most a model may write in one response — blank is automatic (see the response cap).

Beside them are Builder model, the model Synapse AI uses; Model for quick steps, a small, fast model Synapse AI uses to decide what you asked for and whether you approved a plan, and that an agent's helpers run on when their Helper model is the quick-step model (empty follows the builder model; on Cloud, empty uses SynapseOrch's Nova Pro, paid from the model wallet while it holds money, else the builder model); and Embedding model, the model that indexes and searches code repositories, matches each workflow step's semantic cache, and embeds knowledge bases unless Knowledge has its own. The embedding model can be an OpenAI, Gemini, Bedrock, gateway or Ollama one, and each is embedded by its own provider.

See Cloud Providers for per-provider field details, or Local Providers for Ollama / vLLM / LM Studio.

Advanced: the stored model settings

{
  "anthropic_key": "sk-ant-...",
  "model": "claude-opus-5"
}

The provider is chosen from the model name; an older mode setting is ignored.

Knowledge bases have their own defaults — an embedding model, a reranker and a Contextual Retrieval model — under Settings → Models → Knowledge; see Search and settings.

Per-agent model override

In the agent editor (Agents → click an agent), the Model field lets you pick any model that's available on a connected provider. On a self-hosted install, leave it on Use default (…) to follow the default model; on Cloud it shows the workspace's default model until you pick another.

Advanced: equivalent agent JSON

{
  "name": "Fast Router",
  "model": "claude-haiku-4-5"
}

Per-step model override

Inside a workflow, open the configuration panel for an LLM, Tool or Evaluator step and set Model override. Agent steps have no model control in the panel — they use their agent's model, or a model set in the step's JSON. This is powerful for cost management: use a cheap model (e.g. Haiku, GPT-4o-mini) for routing and classification, and reserve capable models for steps that need them.

Advanced: equivalent step JSON

{
  "id": "step-classify",
  "type": "llm",
  "model": "gpt-4o-mini",
  "prompt_template": "Classify: {state.input}"
}

Cost limits

There is no per-run spending cap. To bound what a run can cost, use these levers:

• Per-step model choice — the largest single factor. See per-step overrides.
• Reasoning effort — off or low on steps that have nothing to reason about.
• Max Turns on an agent (on Cloud, Turn limit), which caps how many times the ReAct loop can go round, and Max response length, which caps each reply.
• Step and global timeouts — see Environment variables.
• Prompt caching, which is on by default.

On SynapseOrch Cloud, the plan's credit allowance is a hard bound on every plan: runs stop when it runs out, until you buy more credits or upgrade.

Prompt caching

All providers above transparently benefit from the prompt cache, which is on by default and typically cuts long-conversation cost by 50–80%. On a self-hosted install, Usage's Cache analytics shows the savings per model.