OpenAI Beta
3 min read
OpenAI connects in two ways: with your ChatGPT subscription or with an API key. Each one is a different connection type, and both are created in Settings > Models > OpenAI.
| Connection type | What you pay | Context window |
|---|---|---|
| Subscription | Your ChatGPT plan (Plus, Pro, Business…), with its usage limits. | 872,000 tokens |
| API key | Every token, on your OpenAI Platform account. | 1,050,000 tokens |
Connect your ChatGPT subscription
Create the connection
In Kujira, click Add connection under OpenAI, choose the Subscription type, give it a name that says which account it belongs to, and click Connect.
Authorize Kujira in ChatGPT
Kujira shows you a code. Click Open ChatGPT, sign in with your account and enter the code. When you're done, the connection is saved on its own.
If your account doesn't have sign-in with a code enabled, click Use the browser: you authorize in ChatGPT and paste into Kujira the address it takes you to.
Connect an API key
Create the key
In OpenAI Platform, create an API key. The account needs credit: without it, OpenAI rejects every request and your agents don't answer.
Save it as a connection
In Kujira, click Add connection under OpenAI, choose the API key type, give it a name, paste the key, and save. Use the Test button on the connection to check that OpenAI accepts it.
Turn on an agent
In the agent, open Settings > Model, pick the OpenAI connection under Connection or leave Inherit from the organization, then select an OpenAI model. The change applies on its next reply, with no restart.
What to expect
- The window depends on the connection type, not the model. With the subscription, the model works with 872,000 tokens, which is what ChatGPT allows; with the API key, with 1,050,000. Each conversation's context bar measures against the one that applies.
- With the subscription there's no per-token cost: your ChatGPT plan sets the limit. If you run out, the conversation says so in the chat until it renews.
- With the API key, what OpenAI counts as the model's reasoning is billed as output, and Kujira adds it to the agent's usage. The cost you see in Kujira is an estimate using OpenAI's published rates; the final figure is the one in your OpenAI dashboard.
- With the API key, above roughly 272,000 input tokens OpenAI charges double per token, and Kujira doesn't account for that: in a very long conversation the cost it shows can end up below the real one.