Announcement · Developers team-1 is a new model name on the Atris chat completions API. A fast member answers first, and its words stream right away. A reviewer on a stronger model then checks that answer and either passes it or sends a correction. An app that already uses the OpenAI client changes the base URL and sets the model to `team-1`.

About 0.7 to 1.2 s

First visible word, measured from the caller's side

About 4.3 to 5.8 s

Reviewed answer, start to finish

About 1.5 to 3 s

Time the review adds

$0.75 / $3.50

Price per million input / output tokens

These times come from two small tests with real models. More on that below. ## How it works

You send one request. The fast member starts answering, and you see its words as they arrive. When the fast member finishes, the reviewer reads the answer. If the answer holds up, the reviewer passes it and the call ends. If it finds a mistake, it sends a corrected answer. The correction streams after a separator, so your app can tell the two apart. Very short exchanges skip review. You can also turn review off for a single request. If the reviewer fails or times out, you still get the fast answer. It is marked `unproven`, so you know nobody checked it. ## Try it team-1 works with the OpenAI client you already have. Change the base URL and the model name. ```python from openai import OpenAI client = OpenAI( base_url="https://api.atris.ai/v1", api_key="YOUR_ATRIS_KEY", ) stream = client.chat.completions.create( model="team-1", messages=[{"role": "user", "content": "What is a leap year?"}], stream=True, ) for chunk in stream: if chunk.choices and chunk.choices[0].delta.content: print(chunk.choices[0].delta.content, end="") ``` To turn review off for one request, add `extra_body={"atris": {"review": False}}` to the call. The full reference is at [atris.ai/developers](https://atris.ai/developers#team-1). ## What the stream tells you Every piece of the stream carries one small extra field named `atris`. It says which member is working and whether the answer has been checked. The statuses include `working`, `passed`, `corrected`, `skipped` and `unproven`. Standard clients ignore the field. In our test, the official `openai` Python package read every chunk without changes. The last piece includes a summary of the call: tokens per member, time to first word, total time, the verdict and the price. ## Seat your own agent You can seat a published Atris agent as the fast member or as the reviewer: ```json {"atris": {"members": {"fast": ""}}} ``` The agent answers under its own name. If the agent has its own fee, the fee appears as a separate line. It is never mixed into the token price. ## What we have measured so far We ran a first test on the live API: six calls with real models, after an earlier round of eleven before launch. That is a small test, and there is no benchmark yet. With the built-in fast member, the first visible word arrived in about 0.7 to 1.2 seconds, measured from the caller's side. A reviewed answer finished in about 4.3 to 5 seconds. With a published agent in the fast seat, the first word arrived in about 0.9 to 1.3 seconds once the agent was warm, about the same as the built-in fast member. The first call after a quiet minute can take about a second longer, while the agent's identity is reread. In the earlier round, on one question, the fast member answered that a range of years contained 49 leap years. The reviewer corrected it to 51, which is right. ## Pricing team-1 costs $0.75 per million input tokens and $3.50 per million output tokens, counted across all members. Each call costs at least one cent, so for short calls the cent is the price. ## What we store Atris stores a record of each request: the members, timing, tokens, verdict and cost. That record holds no question text and no answer text. ## What it does not do yet Tool calls through team-1 are not supported yet. Long multi-step jobs are not supported through team-1. Atris has a separate team runs API for long work on a business computer. It is not part of this launch. The review takes time. In our tests it added about 1.5 to 3 seconds to a call. There is no benchmark yet. The numbers on this page come from two small tests. ## Availability team-1 is available today on the Atris API to anyone with an Atris account. Create a key at [atris.ai/dashboard/developer](https://atris.ai/dashboard/developer) and add credits at [atris.ai/credits](https://atris.ai/credits). One credit is one cent.