AI plans and credits
Models and API keys
Add an OpenRouter or DeepSeek key, pick the model each agent uses, and set the task models agents hand harder work to.
Your agents think with an AI model you connect: a plan you already pay for, or an API key. You can connect more than one, and each agent can run on a different model. To connect a ChatGPT, Claude or Google AI plan, see Connect your AI.
Use an API key
- Click your name at the bottom of the left column and choose Settings.
- Choose Models, then click ADVANCED.
- Paste your key into the box beside OpenRouter or DeepSeek and press Enter. The box then shows saved.
On the first screen, Connect your model., you can also click or paste an api key under the three plans and paste a key there. To take a key out, click REMOVE in Settings, then click it again to confirm.
An OpenRouter key also turns on faster browser work: agents hand a run of clicks to a fast model that picks each next click, instead of clicking one turn at a time. Those runs are paid from that key.
Pick the model for an agent
- Under the message box, click the model's name.
- Choose from the Model list. It shows every model your plans and keys offer, with Search models at the top. A model whose limit is used up is marked Usage limit reached.
Beside it, Thinking sets how hard the model thinks before it answers (Thinking unavailable if the model has no levels). The Usage ring shows how much of your plan's limits is left and when they reset. Settings > Usage shows the tokens each agent, routine and model used.
The chat says MODEL LIMIT REACHED
Your plan ran out mid-chat. Click CHOOSE MODEL and pick another model.
The chat says MODEL CONNECTION EXPIRED
Your sign-in has expired. Click OPEN MODELS, sign in again there, and resend your message.
Task models
An agent can hand part of a job to a model picked for that kind of work. Set them in Settings > Models > ADVANCED, under Task models.
| Task model | Used for |
|---|---|
| Fast | Quick lookups |
| Standard | Ordinary research |
| Deep | Hard reasoning, planning and judging |
| Visual | Reading captchas, charts and screenshots |
They come into play when:
- an agent puts a worker on part of its job and names the kind of work (see Hand off big jobs);
- an agent needs a second look at a screenshot or a captcha (Visual);
- an agent is stuck and hands its next turn to a stronger model in the same chat and browser, then goes back to its own model;
- a routine's MODEL is FAST, STANDARD or DEEP instead of THE AGENT'S OWN (see Routines).
Each row starts on Automatic (...), the first model from your Claude or ChatGPT sign-in that fits. Choose another to override it. Only Claude and ChatGPT models are offered here. With neither signed in, the rows say Automatic (none signed in) and the work stays on the agent's own model.