Harbor-1

Harbor-1 is Harbor’s private AI chat. You can research, write, and analyse in one place. Every chat is private by default, and Harbor never trains on a private chat.

Harbor-1 is in closed beta. Join the waitlist on the Harbor-1 page and Harbor will invite you when a place opens.

Models

• Harbor-1 (every plan). Qwen3-4B-Instruct, an Apache-2.0 open-weight model, served by Harbor on rented GPUs.
• Harbor Fast (Pro and Super). The same model on a faster serving tier.
• Claude Opus 5.5 (Super). Anthropic’s model, through Anthropic’s Claude API.

Images are answered by a separate open-weight vision model that Harbor also serves. When a stronger open model passes Harbor’s evaluation, Harbor-1 switches to it, and this page is updated.

How a chat stays private

• In transit. TLS from your device to Harbor.
• At rest. Each chat is encrypted and visible only to your account.
• During inference. Harbor-1 and Harbor Fast run on GPUs Harbor rents and operates. Those prompts do not go to an outside AI lab.
• Claude Opus on Super. When you choose Claude Opus, your messages are sent to Anthropic to generate the reply, under Anthropic’s commercial terms. Choose Harbor-1 for chats that should stay on Harbor’s own serving.
• Training. A private chat is never used to train a model.

Limits and what a reply costs

Every plan has a daily message limit, plus a 7-day and a 30-day allowance. Each reply is measured in tokens and counted against the allowance at that model’s price.

• Claude Opus 5.5 is counted at Anthropic’s list price plus 30%.
• Every turn of a chat resends the conversation so far. On the Claude route, Harbor serves that earlier conversation from a prompt cache instead of sending it fresh. Reused text is counted at Anthropic’s cached rate plus the same 30%, so long chats use far less of your allowance.
• Your billing page shows how much the cache saved you in the last 30 days.

Prompt caching is rolling out on the Claude route and is switched on per environment.

Earn

Earn is a switch on a conversation. It is off until you turn it on, and a new chat always starts private. With Earn on, the conversations you choose to share can be reviewed, and accepted contributions are paid in USDC or Harbor API credits.

The daily check-in draw for Harbor API credits is coming soon. Until it opens, the check-in button says so and no draw runs.

API credits pay for requests made with your Harbor API keys, soonest-expiring first. They are not money and cannot be cashed out.

Plans

• Free, $0. Harbor-1 with a daily message limit.
• Pro, $15 a month. Higher limits and Harbor Fast.
• Super, $59 a month. The highest limits and Claude Opus 5.5.

See Pricing for current prices and limits.

What Harbor never does

• Train on a private chat.
• Use a conversation you did not choose to share.
• Send a Harbor-1 or Harbor Fast chat to an outside AI lab. Only chats where you pick Claude Opus go to Anthropic.