Settings

Three things are yours to decide: which model answers, how it answers, and whether you want to speak and listen. This page brings them together.

Which model answers

Agentic Chat brings several AI models together in one app. You can switch the model per chat at any time, via the model chip in the composer.

Standard models: available with an active subscription, with no credit usage:

• MiniMax M3: fast and versatile.
• GPT-5.6 Luna: affordable and fast for everyday use.
• DeepSeek V4 Pro: strong at code and logic.
• Kimi K2.6: large context, ideal for long documents.

Advanced models: highest quality, credit-based, available in every plan:

• Claude Sonnet 5: strongest Sonnet model with 1M context, especially strong at agents and research.
• GPT-5.6 Sol: creative and powerful.
• Kimi K3: very strong reasoning with a huge context.
• Claude Opus 5: highest quality for complex tasks.

New chats are titled automatically. Below every answer, a note reminds you that the model is an AI and can make mistakes; please double-check important information.

How it answers: custom instructions

Custom instructions let you save your own guidance that is applied automatically to every chat, no matter which model you are using. They are a short, persistent text you set once, added to every conversation as extra context, so you don't have to repeat preferences like form of address or answer style in each chat.

You'll find them in Settings under Account, in the Custom Instructions section (right below your email address). Enter your text, click Save, done.

Examples:

• Form of address: "Call me by my first name."
• Role and context: "I'm a developer based in Vienna."
• Answer style: "Keep answers short and factual."
• Language: "Always reply in English."
• Personal preferences: "Explain technical terms when you use them."

Custom instructions can be up to 2000 characters. You can change the field at any time; if you leave it empty, no instructions are applied.

They apply to all chats and all models. They complement the system guidance but do not replace safety, language, or platform rules, or project context.

Speaking and listening: Voice

Agentic Chat supports voice input and output on the web.

Voice input (transcription)

• Record using the microphone button in the composer.
• Push-to-talk: hold the Shift key to record; when you release it, the audio is transcribed.
• Cmd+Shift+V toggles recording (see Working with Results).
• English is recognized; the transcript appears in the composer.

Voice output (read aloud)

• Have a response read aloud using the speaker or playback icon below it.
• Click again to stop playback.
• Voice selection in the settings: Paul, Oliver, Jane.

The recording format is chosen to be Safari-compatible.

What uses credits

Two settings affect your usage:

• Advanced models are billed through credits; standard models use none.
• Voice is billed via credit hold/settle, so only actual usage is charged.

Both require an active subscription. Details under Credits & Subscriptions.