GPT-6.1 Terra from OpenAI: agent assembles a 3D model in OSWorld 2.1

⚡ The gist in 5 seconds

• OpenAI announced GPT-6.1 Terra and put OSWorld 2.1 at the center of the announcement — a benchmark for AI agents in desktop applications.
• The tasks in it are multi-step: assemble a 3D model from a technical drawing, not answer with text in a chat.
• The description mentions a medium reasoning mode; there are no figures, pricing, or country list in the submitted news item.

🔍 What happened

OpenAI published an "Introducing GPT-6.1 Terra" page at openai.com/index/introducing-gpt-6-1-terra/. The new model is called GPT-6.1 Terra. The main metric the description points to is OSWorld 2.1: it tests how well AI agents operate in desktop applications and carry multi-step tasks through to completion. The example from the text is building a 3D model from a technical drawing.

The second fact from the news item: the measurements are taken "at medium reasoning". The phrase cuts off, so what comes next — percentages, a comparison with other levels, or a caveat about cost — we don't know. How such benchmarks usually work: the agent gets a virtual desktop, a mouse, a keyboard, and a set of applications, and the grader looks at the final state of files and windows, not at the text of the answer. This is general knowledge, not something attached to this specific announcement.

🧠 Our take

Betting on a desktop benchmark is a statement, not an accident. Our assessment: OpenAI is showing that the competition has moved out of chat answers and toward agents that click through menus and save files. You can't assemble a 3D model from a drawing with a single prompt — you need to open a CAD app, figure out the interface, perform a dozen actions in a row, and not lose context. That's exactly what they're selling to enterprise customers.

What this changes for decision-making: models will now be compared not by how pretty their answers are, but by whether they carry a task through to completion in someone else's interface. For your wallet, this means nothing so far — there's no pricing in the news item at all, and no limits either. From experience, agentic modes cost more than regular requests, because a single "answer" is a chain of steps and tool calls. But that's market logic, not a fact from this page.

Who wins — platforms that already sell routine automation: from RPA vendors to agent startups. Who loses — those whose argument was "we're cheaper on chat tasks": a customer with drawings doesn't care about that argument. What bothers me is something else: letting an agent onto a desktop with permission to save and delete files is already an access-rights question, and the news item says nothing about it.

⚠️ What we don't know

As submitted, this is a single line with a truncated description. There's no OSWorld 2.1 result, no comparison with previous models, no price, no date, no country list, and no API details. It's unknown what medium reasoning means in practice — how many steps the model takes before stopping. There's a single source, and you need to verify it yourself: open the page and look at the tables.

Prediction: if the page is official, within a quarter it will feature an OSWorld 2.1 results table broken down by reasoning levels. If there's no breakdown — we'll consider the announcement a showcase without verifiable numbers.

Disclaimer: a single source, the description is truncated — figures and terms are not officially confirmed.

Source: OpenAI