Skip to content

Models and API keys

The chat runs on a language model. There are two ways to power it.

Sign in and start typing — there is nothing to configure. Free turns run on Haiku, and each reply notes how many you have left today. The allowance resets at 00:00 UTC.

Free-tier turns do not include web search, because search bills to the key’s owner and on free turns that is us.

Settings → API keys. Paste a key and the model picker unlocks:

ProviderModels
AnthropicHaiku, Sonnet, Opus, Fable
OpenAIGPT‑5.6 sol, terra, luna
OpenRouterSeveral hundred models from ~45 vendors

With your own key you get unlimited turns, your choice of model, and — on Anthropic and OpenAI — web search. Usage bills directly to your provider account; taker never takes a cut of inference.

Your key is encrypted before storage and cannot be read back by anything, including support. Remove key deletes it outright; chat then falls back to the free tier.

An OpenRouter key reaches many vendors on one account, including models that have no direct integration here.

The picker opens on a short list and searches the rest — start typing a vendor or model name to reach the full catalogue.

Two things are worth knowing about it.

Not every OpenRouter model appears. taker works by having the model call tools: it proposes an order through a tool, and the server prices and places it. A model that cannot call tools cannot do that, and rather than fail strangely — replying “placing that order now” and doing nothing — those models are left out of the picker entirely. Roughly 300 of OpenRouter’s catalogue qualify. OpenRouter’s :batch variants are also excluded: they answer asynchronously, which does not work in a chat window.

Web search is unavailable on OpenRouter. Search is a tool the provider runs on its own side, and its shape differs between Anthropic and OpenAI. There is no equivalent that works across every vendor OpenRouter fronts, so turns on an OpenRouter model have no search.

Auto mode skips the confirmation card for orders under your limit. It requires a verified model.

Verified means the model’s tool calling has been checked as reliable — the frontier models from Anthropic, OpenAI and Google. This is not a judgement of how good a model is in general. It is specifically about whether it holds an exact schema every time, because the failure that matters is not a refusal but a card that looks perfectly normal with one wrong number in it.

Every other model still trades. It simply always shows the card for you to read, the same rule that already applies to any turn that used web search. The picker marks these confirms every trade so you know before you switch.

Haiku is fast and entirely adequate for “long 5x eth 2k”. The larger models are worth it for multi-step requests, questions that need reasoning, or when you want a more capable conversational assistant.

Order construction is deterministic regardless of model — the numbers on a card come from the server, not the model, so a cheaper model does not mean a worse price. It mostly means a less articulate conversation.

Terms of Service · Privacy Policy