GPT-6.1 Terra vs Gemini 4 Argon: Polish Test Results

👁 121 reads

GPT-6.1 Terra costs half as much as Gemini 4 Argon through the API: $1 / $5 versus $2 / $10 per million tokens. But Argon ranks higher on the independent Artificial Analysis index, with 52.6 points to Terra’s 48.1. I tested both models on October 26, 2026, using 12 tasks in Polish, office tasks from OpenAI’s announcement, technical drawings from a scan and a photograph, screenshot-based form filling, and 20 questions about Polish regulations. Terra either won or tied in every category, while Argon failed to finish its answer twice.

Price, context window and input types

Terra’s price comes from OpenAI’s pricing page; Argon’s price, context window and input types come from OpenRouter’s model list; and the benchmark scores are from Artificial Analysis. I checked all three on October 26, 2026. Argon accepts audio and video inputs; Terra does not. In the app, Terra is available only in ChatGPT Work and Codex (Plus, Pro, Enterprise and Edu plans). For details on accessing Argon in Poland, see my Gemini 4 Argon guide.

12 tasks in Polish: 65 to 64

I used the same set of tasks as in my GPT-6.1 Terra guide: 12 tasks worth 0-6 points each, some scored automatically and others by blind AI judges. Gemini 4 Argon alone scored Terra’s answers, while GPT-6.1 Sol alone scored Argon’s, so neither model judged its own family. Argon answered on October 24, Terra on October 26, and I scored both sets in a single batch on October 26. The test prompts and responses were in Polish.

Correction: my first summary gave Terra 69/72. The automatic scorer accepted its logic puzzle response, even though it gave three options instead of one: “B: 2, 3, or 4.” On manual review, that answer earned 2 out of 6 points, bringing Terra’s total to 65/72.

The 12 tasks cost $0.0641 with Terra and $0.1280 with Argon. They took 85 seconds and 100 seconds, respectively, with requests run in parallel; treat the times as approximate. I consider a 1-point difference a tie: rescoring the same set can shift the results by 0-2 points.

Office tasks from OpenAI’s announcement

OpenAI highlights Terra’s ability to handle technical drawings, quotes and forms, so I gave Argon the same four tasks. Both models scored 12/12 on a circuit-board drawing, 13/13 on a flange drawing and 16/16 on a 16-field form. The difference came down to a sales quote worth 10 points: Terra scored 9, Argon 6.

The brief said, in a translation of the Polish original: “Free delivery on orders over 1,000 original Polish currency units, excluding tax.” Argon’s quote said free delivery applied only in Lublin, a restriction the brief never mentioned. The judge, GPT-6.1 Sol, said: “The free delivery offer was unjustifiably restricted to Lublin, changing the terms of the offer.” That is the worst kind of error in a customer letter because it sounds plausible. Both models correctly calculated the amounts: 1,564.00 original Polish currency units before tax, 1,923.72 original Polish currency units including tax, and 1,827.53 original Polish currency units after the discount. The four tasks cost $0.0134 with Terra and $0.0256 with Argon.

Technical drawings from a scan and a photo: a tie

Both models scored perfectly on clean, computer-generated drawings, so I made the test harder. I used a flange drawing as a noisy, skewed JPEG 30 scan and as a photo of a printed copy taken at an angle. I also made a new shaft drawing with h6 and k6 tolerances, an M20x1.5 thread, a keyway and Ra 0.8 surface roughness, then tested it both clean and as a photograph.

On the scan, both models made the same mistake: they read the bolt circle as “Ø60” instead of Ø50 and calculated the spacing between holes as 30 mm instead of 25 mm. I’ll admit that the 5 does look like a 6 in this scan. Argon responded faster, taking 6-11 seconds per image compared with Terra’s 9-14 seconds.

Filling in a form from screenshots

In my GPT-6.1 Terra computer-use test, the models filled out a two-page leasing application using only screenshots. Terra filled in all 15 fields in all four attempts, taking 19 steps and costing about $0.047. With the same instructions, which specified pixel coordinates, Argon clicked beside the fields and had not submitted the form after 40 steps. Two attempts cost $1.02. After I changed the instructions to use a 0-1000 coordinate scale, Argon matched Terra: 15/15 fields, 19 steps, $0.082 and 72 seconds. Both models can do the task, but Argon needs a different coordinate format.

Polish regulations: Argon did not finish its answers

In a test of 20 questions about Polish regulations (16 about real laws and four about laws that do not exist), Terra scored 20/20 at all three reasoning levels. In two attempts, Argon stopped at the same point, after question seven. OpenRouter returned an error status from Google AI Studio. The seven answers it did provide were correct and quoted the regulations almost word for word. Google did not give a reason, so I won’t speculate. If you generate long responses quoting legislation, check that the answer is complete.

Which model should you choose?

• Terra - For Polish letters and quotes, forms, and reading documents and drawings at scale. It matched or beat Argon in every test I ran, at roughly half the price.
• Argon - If you need audio or video input, which Terra does not accept, or a higher score on the most difficult tasks in Artificial Analysis. For more of my results, see Gemini 4 Argon: 23-task test.

Frequently asked questions

Is GPT-6.1 Terra better than Gemini 4 Argon?

On my office tasks and Polish-language prompts, Terra either won or tied: 65 to 64 on the Polish tasks and 50 to 47 on the tasks from OpenAI’s announcement. Argon scored higher on Artificial Analysis, with 52.6 points to Terra’s 48.1.

Which model is cheaper?

Terra costs $1 / $5 per million tokens, compared with Argon’s $2 / $10. My 12 Polish-language tasks cost $0.064 with Terra and $0.128 with Argon.

Which model reads technical drawings better?

They tied at 57/58 on scans and photos, making the same reading error on the most difficult scan. Argon responded faster; Terra cost less.

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

Sources and test date

• OpenAI: Introducing GPT-6.1 Terra - price, availability and use cases (checked October 26, 2026)
• OpenRouter: API model list - Gemini 4 Argon’s $2 / $10 pricing, context window and input types (checked October 26, 2026)
• Artificial Analysis: GPT-6.1 Terra - score of 48.1 and comparison with Argon’s 52.6 (checked October 26, 2026)
• My test: OpenRouter (Terra - OpenAI endpoint; Argon - Google AI Studio), medium reasoning, October 24-26, 2026

Kinetyra