Models
Every step an employee takes, from planning a task to writing a reply, runs on a language model. Orqane gives you access to models from several providers, lets you pick one per employee or per skill, and can route each step automatically.
Available models
Section titled “Available models”Orqane offers model families from the main providers. The exact list in your workspace is in Settings → Models.
| Provider | Families | Strengths |
|---|---|---|
| Anthropic | Claude Opus, Sonnet, Haiku | Reasoning, tool use, writing, computer use |
| OpenAI | GPT, o-series | Reasoning, structured output |
| Gemini Pro, Flash | Long context, multimodal input | |
| Mistral | Mistral Large, Medium, Small | EU hosting, multilingual |
| Open-weight | Llama, Qwen, and others hosted by Orqane | Low cost, EU hosting |
Orqane sorts models into three tiers to make choosing simpler: Frontier for the hardest reasoning, Balanced for most work, and Fast for short, routine steps.
Choose a model
Section titled “Choose a model”You set models at three levels. The most specific one wins.
| Level | Where | Example |
|---|---|---|
| Workspace default | Settings → Models | Automatic |
| Employee | Employee profile → Model | A support employee on a Balanced model |
| Skill | Skill settings → Model | The “Draft a contract” skill always on a Frontier model |
Automatic routing
Section titled “Automatic routing”With Automatic, Orqane picks a model for each step based on what the step needs. Planning a complex task goes to a stronger model; classifying an email or extracting a field goes to a faster, cheaper one.
Choose what to optimize for:
| Preference | Effect |
|---|---|
| Quality | Prefers Frontier models. Higher cost, best results on hard tasks. |
| Balanced | Uses Frontier models for planning and hard steps, Fast models for routine ones. Default. |
| Cost | Prefers Fast models, and moves up only when a step fails or a check flags low confidence. |
| Latency | Prefers the fastest models. Useful on voice and live chat. |
You can also limit routing to an allowed list of models, for example only EU-hosted ones. Each step in Activity shows the model that ran it, the tokens used and the cost.
Fallback
Section titled “Fallback”If a provider is down, slow or rate limited, Orqane retries the step on the next model in the fallback chain. By default, the chain stays in the same tier and moves to another provider. Set your own chain in Settings → Models → Fallback.
Fallback respects your restrictions: if an employee may only use EU-hosted models, it only falls back to EU-hosted models. If no allowed model is available, the run pauses at a checkpoint and resumes when one is.
EU-hosted models
Section titled “EU-hosted models”Workspaces in the EU region can require that all model calls stay in the EU. Turn on EU-hosted models only in Settings → Models. Orqane then uses only models served from data centers in the European Union, including Mistral models and open-weight models that Orqane hosts.
Bring your own key
Section titled “Bring your own key”To use your own contract with a provider, add your API key in Settings → Models → Provider keys. You can add keys for Anthropic, OpenAI, Google, Mistral, Azure OpenAI, Amazon Bedrock and Google Vertex AI.
| Orqane-provided models | Your own key | |
|---|---|---|
| Model usage billed by | Orqane, in credits | The provider, on your account |
| Platform fee | Included | Charged in credits per step |
| Data processing terms | Orqane’s | Yours with the provider |
| Rate limits | Orqane’s | Your account’s |
Keys are encrypted and never shown again after you save them. If a key fails, Orqane falls back to Orqane-provided models only if you allow it in Fallback.
Credits and budgets
Section titled “Credits and budgets”Model usage is billed in credits. Each model has a price per million input and output tokens, shown in credits in Settings → Models. Credits spent by an employee count toward its budget. When an employee nears its budget, Orqane can switch it to cheaper models before stopping it. Set this in the employee’s Budget tab.