Frequently asked questions

Cloud models — Claude, GPT, Gemini, Grok, Kimi, DeepSeek, GLM, MiniMax and hundreds more through OpenRouter and xAI. Open models — Qwen, GPT-OSS, Devstral, DeepSeek Coder and R1, Gemma, Kimi-VL and others: they run on our own servers and count against your limit at low prices.

Open models run on our own GPU. They are the cheapest option: they count against your limit at low market prices ($0.02–0.40 per million input tokens depending on size), and your messages are not sent to third parties. Cloud flagships are stronger at hard tasks, but every request costs money — that's why they have limits.

Yes. Press + next to the model name and add up to four models (depending on your plan). In “Take turns” they answer one after another and see each other's replies; in “Side by side” each answers on its own, in columns, so you can compare. Each answer is charged at its model's price.

Yes, from the Go plan. Create an API key in Settings → API & coding agents and copy the ready-made config for Codex CLI, Claude Code, opencode, Cline, Roo, aider or any OpenAI-compatible app. The agent then works with your project files on your own computer, and usage counts against your plan limit, the same as chat.

Turn on Agent under the message box and describe a task. The model makes a plan, works through it with its tools — web search, a code sandbox, files — checks its own work and reports back. It takes longer and uses more of your limit than a normal answer.

Reasoning models think before they answer. Those internal tokens are billed by the provider like regular output. Higher levels (Low → Max) give more accurate answers on hard problems but cost more and take longer. For simple questions use None or Low.

Some providers (for example OpenAI for o-series models) don't return the reasoning text — only its size. In that case the answer shows how long the model thought and how many tokens it used.

No. Modellaris is only the interface. Every model is made by its own developer (Anthropic, OpenAI, Google, Amazon, Alibaba and others), and we tell the model so. If a model still gets it wrong, that's a hallucination — please let us know.

Temperature controls how varied the answers are: low is more precise and consistent, high is more varied. You'll notice it if you regenerate the same answer several times. Reasoning models (o-series, GPT-5/6 and similar) ignore temperature because the provider fixes it; adjust the reasoning depth instead.

Yes, every model. If a model can't see, an open vision model on our server describes the image for it. Video is split into key frames and an audio track, and speech is transcribed.

Models with tools decide when to search: web, Wikipedia, Reddit, Hugging Face, OpenRouter, GitHub, arXiv, Stack Overflow. Higher plans get deep search, which reads full pages. Source links appear right in the answer.

In Settings → Agents you create assistants with a name, a model, a role (for example Planner, Coder, Reviewer) and tools. Add several to a chat and give them a task: they talk to each other in real time, you see every message, and you can step in at any moment. Agents stop when the work is done.

Turn on Research under the message box. The model plans the search, reads dozens of pages, checks facts across sources and writes a report with links to every source as a file. It takes a few minutes and uses more of your limit than a normal answer.

Yes. Open a project, upload or create files, and press Run on an HTML file: the page opens in a window right on the site. It runs in an isolated sandbox with no access to your account or other pages.

A project is a folder with your files and instructions. Upload files or a whole .zip; every chat in the project and every model can read, create and edit them. You can download everything as one .zip at any time.

Only what you approve. When the model notices a lasting preference, it shows a card “Save to memory?” with the text. Nothing is saved until you press Save. You can edit or delete memory in Settings.

Yes. Press the headphones button for a hands-free conversation: you speak, the model answers aloud, and you can interrupt it. You can also dictate a single message or listen to any answer.

Yes. Ask the model to “give it as a file” — it creates the file (code, document, table) and it appears as a card with a Download button. Several files can be downloaded as one .zip.

The beginning of the conversation is automatically summarized so the model stays within its context window. All your messages stay visible; you can also compress a chat manually from its menu.

No. Images, music, voiceover and 3D are available; models understand video but don't generate it.

For cloud models — in dollars, at the actual price of each request as reported by the provider. Limits renew on a rolling basis: every 5 hours and every week. You can never go negative: if the remaining limit isn't enough for an expensive model, we suggest a cheaper one. Click the price under an answer to see the breakdown: input, output, reasoning.

Usually because of an expensive reasoning model. For example, o3 Pro costs $20 per million input tokens and $80 per million output tokens, and reasoning counts as output. Two or three thousand reasoning tokens already cost several cents.

One-time top-ups of $3.99, $9.99 or $19.99 that you can use on top of any plan, including Free. Credits are spent only when your 5-hour or weekly limit is used up (or your daily quota of images, music and 3D), at the real price of the model. With credits on your balance, Free also opens Go-level models. Credits do not expire.

$20 per member per month, from 2 members. Every member gets Plus-level limits, and the team shares projects (with files) and agents. Chats stay private to each person. The owner invites people by email or link and adds or removes seats at any time; billing is one invoice.

Payments are handled by our payment provider, Creem, which acts as the seller of record. Your card details are entered on their secure page and never reach us; they also issue receipts and handle taxes.

Anytime, in account settings or by emailing support. You keep access until the end of the paid period. Refund terms are on the Payments & Refunds page.

Sexual or nude content, anything sexual involving minors, deepfakes of real people, graphic violence, hate, malware, fraud and other illegal use. Requests are checked automatically. Repeated violations lead to suspension without refund. Full rules: Acceptable Use Policy.

Automatic checks sometimes make mistakes. Rephrase the request; if you believe it was a mistake, write to us from your account email and we will review it.

No. We don't train models on your messages and we don't sell your data. Cloud providers receive the request text in order to answer it; their policies are described in our Privacy Policy.

Settings → Privacy: “Delete all chats” or “Download my data”. To delete everything, use Settings → Account → “Delete account”. For a single conversation, use a temporary chat.