Model provider
Choose who processes your organization's model requests, and what happens to them afterward.
The model provider decides who processes your organization's requests and what happens to them afterward, which is what a compliance or IT policy usually turns on. Admins set it: open the user menu, choose Workspace settings, and set Inference provider. The choice applies to every member and every deal.
The choice
| Provider | What it means for your data |
|---|---|
| Amazon Bedrock | Requests are processed within AWS infrastructure. |
| Claude Platform on AWS | Requests are processed by Anthropic outside the AWS boundary and retained for up to 30 days. |
| Your own API keys | Requests go to your firm's own Anthropic, OpenAI and xAI accounts, under those vendors' terms. |
If your firm's policy requires deal data to stay within AWS, choose Amazon Bedrock. Where that constraint doesn't apply, Claude Platform on AWS is open to you as well, at the cost of Anthropic processing the requests and keeping them for up to 30 days.
Amazon Bedrock offers Fable, Opus, Sonnet, Sol, Terra, Astra, and Grok. Claude Platform on AWS offers Fable, Opus, and Sonnet.
Your own API keys
Choose this when your firm already holds vendor accounts and wants model usage billed there. Enter a key for each vendor you use; an optional base URL points a vendor's requests at a compatible proxy of your own. Each key is checked with its vendor when you save, and a rejected key keeps the previous settings.
Only the models of the vendors you entered keys for are offered: Anthropic for Fable, Opus and Sonnet; OpenAI for Sol, Terra and Astra; xAI for Grok. The vendors bill you directly for that usage, and Cap Orbit charges only its margin on it instead of the full per-token rate. Document text extraction for scanned PDFs stays on Amazon Bedrock at the standard rate. Keys are stored encrypted and are never shown again after saving; enter a new key to replace one, or tick Remove to drop it.
Speed tiers
Under Speed tiers, tick the tiers sessions may use: Anthropic Fast for Opus, OpenAI Fast for Sol, Terra, and Astra, and OpenAI Ultrafast for Astra. Each ticked tier appears as its own row in /model, where a member picks it for that session only; every session starts at standard speed. The vendor bills a tier at its own higher token price, two times the standard rate for Fast and six times for Ultrafast, and Cap Orbit's margin follows that price.
On Amazon Bedrock the same block offers one tier, OpenAI Ultrafast on Astra, billed at six times the standard Bedrock rate. Claude Platform on AWS offers no speed tiers.
When the change takes effect
The choice applies to newly opened sessions only. Sessions already open keep the provider they started with; open a new session to pick up the change.
Learn more
- Sessions: how to open a new session so the change applies.