GPT-6.1 Terra: what it is, pricing and how to use it

GPT-6.1 Terra is OpenAI's new mid-tier model. It reads screens, clicks through apps and writes, at about half the price of the model it replaces. In the API it costs $1 per million input tokens and $5 per million output tokens, and it's rolling out now in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users.

The posts that announced it

@OpenAI announced it today as "delivering fast and capable computer use at a lower price." The post sells Terra as a model for your whole workday, one that understands what's on screen, acts across apps and writes more naturally than GPT-5.6 Terra did.

A few minutes later the benchmarking firm @ArtificialAnlys posted its own numbers, saying Terra is "pushing the cost efficiency frontier." It's the first Terra-class model since GPT-5.6, and by Artificial Analysis's count it gains 6 points on their Intelligence Index while costing about $0.40 per task at max effort, down from $1.40. The number I care about more: on their AA-Omniscience test, Terra guesses on 38% of the questions it can't answer. GPT-5.6 Terra guessed on 88%.

If the lineup has confused you, 9to5Mac's launch write-up explains it. Terra got skipped when GPT-6 Sol and Luna shipped last month. From cheapest to most capable, the order is now Luna, Terra, Sol, Astra.

GPT-6.1 Terra pricing

From OpenAI's model page:

That's half the list price of GPT-6.1 Sol, which charges $2 and $10. Per task, the gap is wider. Nerdschalk reproduced OpenAI's chart for Agents' Last Exam, a test of agents using work software. At medium effort Terra scores 57.6% for about $0.29 a task, the same score GPT-6 Astra gets at $4.82. At max effort Claude Opus 5.5 still leads, 63.2% at $8.82 a task, against Terra's 60.0% at $1.01.

Is GPT-6.1 Terra free? No. The launch covers Plus, Pro, Enterprise and Edu, and the API's free tier doesn't support it.

How to use GPT-6.1 Terra

Nerdschalk's walkthrough of OpenAI's docs covers each place you can pick it:

• In ChatGPT Work on the web, open the model and reasoning control under the message box and choose GPT-6.1 Terra. The desktop app has the same control, with the full list under Advanced.
• In the Codex CLI, update to 0.159.0 or later and run codex -m gpt-6.1-terra, or switch mid-session with /model. To make it your default, set model = "gpt-6.1-terra" in config.toml.
• In the API, call gpt-6.1-terra. Tool calling, computer use included, needs the Responses API. Reasoning effort goes from none to max, and medium is the default.

OpenAI's example jobs tell you where it's aimed: filling out forms from source documents, reading technical drawings, drafting proposals, matching catalog entries to product photos, editing videos. For the hardest coding and in-depth research, OpenAI still points you to Sol or Astra.

Here's a first prompt to try in ChatGPT Work: "Open this supplier invoice PDF, pull the vendor, date, line items and total, and fill them into the bill form in my accounting app. Stop before saving and show me what you entered." It tests screen reading and form filling, and it puts a checkpoint in front of the save button.

My read

Everyone will lead with the price. I'd watch the guess rate. An agent working inside your apps does its damage when it invents a field value and keeps going. Going from 88% to 38% on Artificial Analysis's test doesn't make Terra trustworthy, but it does mean you can check the fields that matter instead of every one.

Two cautions. Most of the benchmark numbers come from OpenAI grading its own model. And Artificial Analysis found only the low and medium effort settings are good value, so turning the dial to max mostly gets you a bigger bill. OpenAI also rates Terra at its Critical cybersecurity threshold and ships it with Astra's safeguards. Put that next to the AGENTS.md Downloads attack and the usual agent rule applies: limit what it can open and what it can send.

What this means for a business adopting AI

Computer use has been the slow, expensive end of AI. Terra, along with tools like Cursor Computer Use, makes it an everyday option for dull screen work like rekeying invoices, copying leads from email into a CRM, or filling out the same portal form every week. Pick one of those jobs and run it at medium effort with a person approving the last click. Track the errors for a couple of weeks before you give it more rope.

If you want help picking which task to hand an agent first, New Face Design runs a free process audit. We map where your team's hours go and flag the steps a model like Terra could take off their plate.