GPT-6.1 Terra Computer Use: Form-Filling Test

👁 129 reads

OpenAI describes GPT-6.1 Terra as a model with “precise computer use” - controlling a computer from screenshots - and lists form filling as one possible use. I tested that claim with a two-page leasing application: the model saw only a screenshot and a customer email containing traps, while the browser carried out each action, whether clicking or typing. Terra completed the form without errors in all four attempts, using 19 steps each time, for about $0.047 per attempt and in under a minute. The test form and email were in Polish, so the results are from a Polish-language test.

What OpenAI promises

In its announcement on 26 October 2026, OpenAI describes Terra as a fast model combining accurate visual reasoning with precise computer control. Its examples include filling in forms; in the API, it also mentions reading invoices and entering data into accounting software. On the OSWorld 2.1 test, which involves tasks on a real operating system, Terra reportedly exceeds the best GPT-5.6 Terra score at the medium reasoning level. OpenAI shows the figures only in a chart.

In the API, computer control is provided by the computer_use tool, which is available only through the Responses API. I tested through OpenRouter, so I built a simpler loop: the model receives a screenshot and returns one action as JSON; a Playwright script carries it out and takes another screenshot. This is less than OpenAI’s full tool, but it tests the same basic abilities: whether the model can target fields and interpret the data correctly.

The form and the email with traps

I created the form myself. Page 1 asked for personal details and an address; page 2 asked for the amount, lease term, item and consents. The region and lease-term lists were custom dropdown menus, not standard select fields, so the model had to click to open them and then select an option. The fields had validation: the phone number had to contain 9 digits with no spaces, and the tax ID had to contain 10 digits with no hyphens.

The model took the information from a customer email written in Polish. The following are English translations of the original Polish test material, not prompts retested in English:

• “Phone: 601 234 567 - actually, sorry, that’s my work number (…) please use my personal number: 512 900 118.” The form needed 512900118.
• “Tax ID 521-301-72-28.” The form needed the number without hyphens.
• “A cutting plotter (…) a three-year lease.” The item was a machine/equipment, and the term was 36 months.
• “Please don’t send me marketing offers.” The marketing consent had to remain unchecked, while the required data-protection consent had to be checked.
• The email did not specify a region. The model had to infer it from postal code 20-346 and the city of Lublin.

The instruction to the model was only a few sentences. Its response had to be a single JSON object in one of these formats. These are English translations of the action formats used in the Polish test, not a new test:

{"action":"click","x":123,"y":456}
{"action":"type","x":123,"y":456,
 "text":"..."}
{"action":"scroll","dy":300}
{"action":"done"}

How Terra worked: 19 steps

Terra’s first attempt used the medium reasoning level. Here are its actions in order, with coordinates in pixels on a 1280 x 800 screen. The accompanying text is translated from the Polish test record:

1. type (412,140) "Tomasz"
2. type (792,140) "Wierzbicki-Nowak"
3. type (408,203) "88030312342"
4. type (789,202) "5213017228"
5. type (420,277) "512900118"
...
10. click (810,417) region list
11. click (690,501) "Lublin region"
12. click (301,461) "Next"
13. type (401,140) "24600"
14-15. click term list, then "36 months"
16-18. click "machine/equipment",
             data-protection consent, "Submit"
19. done

Terra made no unnecessary moves: it filled every field on the first try, never triggered validation and did not check the marketing consent box. It entered the personal phone number, omitted the hyphens from the tax ID and selected the region itself. Each step took an average of 2.6 seconds, with 1,700-2,060 input tokens per step (the screenshot plus action history). The next three attempts were identical, including one at the low reasoning level: the same number of steps and the same result.

One caveat: every model copied the email address exactly as it appeared in the email, in uppercase: “T.Wierzbicki@Studio-Lipa.pl”. The form allowed this, so it was not an error. But if your system treats uppercase and lowercase letters as different, say so explicitly in the instruction.

Results for all models

• Sol did exactly what Terra did, in the same number of steps, but cost twice as much.
• Luna selected the wrong region in one attempt, then reopened the list and corrected its choice, which brought that run to 21 steps. It cost 10 times less than Terra.
• Gemini 4 Argon, when instructed to use pixel coordinates, clicked beside the fields. It entered the first name in three different places and still had not submitted the form after 40 steps. This was not a limitation of the model itself: Argon returns coordinates on a 0-1000 scale. After I changed one sentence in the instruction and converted the scale, it completed the form without errors. That lesson cost me about $1.

What this means in practice

A simple form, even with custom dropdown menus and validation, was a routine task for Terra. Its result matched the more expensive Sol while costing half as much. For 1,000 such applications, the model alone would cost about $47, compared with about $94 for Sol and about $5 for Luna.

This test did not cover real websites with ads, cookie pop-ups or CAPTCHAs; long tasks across multiple websites; or the full computer_use tool in the Responses API. In ChatGPT, Work mode handles similar tasks; I covered it in ChatGPT Work: What It Is. I also covered how OpenAI makes computer use available to developers in OpenAI’s Agents API. Other Terra results, including a test conducted in Polish, are in the GPT-6.1 Terra guide.

Frequently asked questions

Can GPT-6.1 Terra control a computer?

Yes. It is one of the main capabilities OpenAI has announced. In my test, it filled all 15 fields of a two-page form in all four attempts, without a single misclick.

How much does computer use with Terra cost?

You pay for tokens: each screenshot with the action history uses about 2,000 input tokens per step. My 19-step form cost $0.046-$0.048 per attempt.

Terra or Luna for automating forms?

Luna also scored 15/15 on this simple form, at about $0.005, but it selected the wrong option once and had to correct it. Terra was faster and needed no corrections.

Does computer use work through OpenRouter?

Sending a screenshot and receiving an action as JSON works through the standard chat API; that is how I tested it. OpenAI’s computer_use tool requires the Responses API.

Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →

Sources and test date

• OpenAI: Introducing GPT-6.1 Terra - computer use, forms and OSWorld 2.1 (accessed 26 October 2026)
• OpenAI API: GPT-6.1 Terra model - the computer_use tool and Responses API requirement (accessed 26 October 2026)
• My test: custom HTML form, Playwright (Chromium, 1280 x 800), OpenRouter, medium reasoning (one Terra attempt at low), 26 October 2026, about 21:05-21:30; 11 attempts, total cost $1.49.

Kinetyra