Tartarus AI features
Everything the product does, and the reason each part works the way it does. If you only want the summary, the home page has it in eight lines.
Answers stream, and stop when you say stop
Replies arrive token by token over a server-sent event stream, so you start reading a long answer while the rest is still being written rather than watching a spinner for a minute.
The stop control is a real cancel. It tells the server to abandon the generation instead of only closing the browser's end of the connection, which means a run you abandoned stops consuming your allowance and stops holding a slot other people are queued behind.
Reasoning arrives on its own channel
Reasoning models work through a problem before they answer. Tartarus AI streams that thinking on a separate channel and renders it in a collapsible panel above the reply.
It is never spliced into the answer text. The paragraph you copy out is the answer, with none of the model's working left in it, and the working is still there if you want to check how it got there.
Web search with the sources shown
When a question depends on current information, the model runs a search and reads the pages it finds. Every source it used is listed above the reply, so a claim about this week comes with the trail that produced it rather than on trust alone.
Page fetching resolves the target host before connecting and refuses private and link-local addresses. That guard matters because the model chooses the URL and the model's input is ultimately your text: without it, a crafted prompt could reach internal infrastructure.
Branching instead of overwriting
Edit one of your messages, or regenerate a reply, and the previous version is kept as a sibling rather than being replaced. A switcher on the message moves between them.
The practical effect is that trying a second phrasing costs nothing. You can explore an alternative and come back to the original without having saved it somewhere first.
A system prompt per conversation
Every conversation carries its own system prompt, up to 4,000 characters, set on the conversation rather than on the account. A thread can be told what voice to write in, what format to return, or what it is allowed to assume, and the next thread starts clean.
On a platform whose whole point is that it does not add a second layer of restrictions, the layer you do control is worth having. Nothing is prepended behind it that you cannot see.
Search is a loop, not one lookup
A turn that needs current information can run up to three search-read-answer rounds: the model searches, reads what it found, and decides whether it now knows enough or needs a different query.
On the last round the tools are withheld and the model has to answer from what it gathered. That is deliberate — it is what stops a turn ending in an empty bubble because the model kept searching until it ran out of room.
History you can actually find things in
Conversations are searchable by title. Ones you return to can be pinned to the top of the list, and every conversation can be renamed.
Deleting is a soft delete, so a conversation removed by accident is recoverable rather than gone.
An OpenAI-compatible API and CLI
Pro and Ultra plans mint API tokens for an OpenAI-compatible endpoint. Existing clients, SDKs, editor integrations and agent frameworks work against it without a code change: set the base URL and the key, and the requests you already send keep working.
Plan access is re-checked on every request rather than only at the moment a token is minted, so a token cannot outlive the plan that created it.
Code and markdown rendered properly
Replies render as markdown with syntax-highlighted code blocks, copy buttons, tables and lists. Model output is sanitized before it is inserted into the page, so a reply cannot inject anything into the app.
Flat limits, not a meter
Every plan carries a weekly token budget and a session budget, and nothing is billed per request. There is no per-message cost to weigh up and no bill that grows while you work.
Each plan also has a context ceiling and a per-reply output cap. In the browser, older turns are trimmed automatically to fit the ceiling rather than failing the request; over the API an oversized request is rejected so your client can compact and retry.
Crypto billing, no card on file
Plans are prepaid in USDT on TRC20 or ERC20, Bitcoin, Ethereum, Litecoin, Tron or Monero. There is no card stored, no billing profile and no recurring charge, so there is also no subscription to remember to cancel.
Payment confirmation is idempotent and replay-guarded: a webhook delivered twice credits the plan once.