GPT-6.1 Terra vs Claude Sonnet 5.5: Polish Test Results

Update (9 Nov 2026): Claude Fable 5.2 and Mythos 5.2 are out - prices, access and benchmarks are in my Claude Fable 5.2 guide.

👁 114 reads

GPT-6.1 Terra and Claude Sonnet 5.5 are mid-tier models from OpenAI and Anthropic. Terra’s API rates are exactly half Sonnet’s: $1 / $5 versus $2 / $10 per million tokens. On October 26, 2026, I gave both models the same set of tasks: 12 tasks in Polish, four office tasks involving technical drawings, a quote and a form, 20 questions about regulations, and the same letter to a tax authority 10 times. They tied on the Polish-language test, scoring 65 to 64.5 points. Sonnet gave the wrong answer in every trial of the tax-letter task.

API pricing and availability

Terra’s prices are from OpenAI’s price list. Sonnet 5.5’s prices and both context-window figures are from OpenRouter’s model list, checked October 26, 2026. Terra is not available in the standard ChatGPT chat; it is available in Work and Codex on Plus, Pro, Enterprise and Edu plans. See my full GPT-6.1 Terra pricing guide for the 272,000-token threshold and Fast mode.

Polish-language test: 65 to 64.5

This was the same set of tasks I described in my GPT-6.1 Terra guide: a letter to an authority, idioms, a summary with a word limit, a translation with false friends, a VAT calculation, date formats, a Polish national ID code, a logic puzzle, JSON, factual questions, information retrieval from 28,000 tokens and proofreading. Each task was worth 0-6 points. Some scores were calculated automatically; others were assigned blindly by two AI judges, GPT-6.1 Sol and Gemini 4 Argon. Gemini alone judged OpenAI model responses. Both models ran through OpenRouter with medium reasoning. We tested the prompts in Polish; the examples and descriptions here are translations of that Polish test material, not new English-language tests.

Correction: in my initial summary on October 26, Terra had 69/72. When I added Sonnet, I manually checked the logic puzzle and found that Terra had given three options instead of one: “B: 2, 3 or 4.” The automated scorer counted that as correct. Since the puzzle has one solution, Terra gets 2 out of 6 points, bringing its score to 65/72.

The 12 tasks cost $0.0641 on Terra, which used 5,873 output tokens, and $0.1678 on Sonnet, which used 8,099 output tokens. They took 85 and 65 seconds, respectively, with queries run in parallel; timings are approximate. The same judges scored Sonnet’s answers in a new batch alongside 11 earlier models. Terra got the same score in that batch as it did on October 26, so the results are comparable. I consider the 0.5-point difference a tie.

One example from the translation task: Sonnet translated the false friend “sympathetic” as “the new manager is quite likable,” although the intended English meaning was “understanding.” Terra avoided the trap. In proofreading, Sonnet changed a Polish expression meaning “with empty hands” to an alternative form and listed the change as a correction, even though both forms are correct. These examples are translations of the Polish test material.

Office tasks: drawings, a quote and a form

Terra had received the same four tasks on its launch day: two technical drawings to interpret, including a weight calculation; a printing quote for a clinic with VAT and a discount; and a form with 16 fields to fill from a client email containing traps. We tested both models in Polish.

• Plate weight. Sonnet read all the dimensions but gave a weight of 333.5 g. The correct result was 408.2 g, using a steel density of 7.85 g/cm³ and dimensions of 120 x 80 x 6 mm minus five holes. Terra calculated it correctly.
• English loanword in the quote. The Polish test instruction asked for a quote without English loanwords. Sonnet left “Roll-up 85×200 cm with bag” in the table; Terra used the Polish equivalent, “roll-up advertising stands.” Sonnet did end its letter with a specific request for a reply, while Terra used the formula “I remain at your disposal.” The judges deducted one point from Terra for that closing. The quoted examples are translated from the original Polish test.

Legal questions: Sonnet cited the wrong provision every time

The first task in the set asked for a letter requesting payment of overdue tax in installments. Under the Polish Tax Ordinance, the relevant basis is Article 67a § 1, point 2. Point 1 concerns deferring tax payments and paying tax in installments before the tax is overdue; point 3 concerns remission. I ran the same prompt repeatedly. The task and its legal reference concern Polish law, and the test was conducted in Polish.

Each time, Sonnet wrote a sentence equivalent to: “Under Article 67a § 1, point 1 of the Act of August 29, 1997 - Tax Ordinance, I request that the overdue amount be paid in six monthly installments.” This is a translation of Sonnet’s Polish test response. Sonnet made the same mistake in a separate test of 20 questions about regulations, where it scored 19/20 while Terra scored 20/20. Sonnet’s error is consistent, which makes it easier to spot, but also easier to miss if you have checked a correct version just once.

Which model should you choose?

• Terra - choose it if you write Polish letters, quotes and emails, interpret drawings and forms, or have a high volume of tasks. It matched Sonnet’s Polish-language score at 2.6 times lower cost and cost 2.9 times less on the office tasks.
• Sonnet 5.5 - choose it if you value code with comments and clear answers to puzzles. Sonnet gave one correct answer to the logic puzzle, while Terra gave three options.
• With either model, check legal references in letters. Both models confidently cited the wrong provision.

For a comparison with Anthropic’s more expensive model, see GPT-6.1 Terra vs Claude Opus 5.5.

Frequently asked questions

Is GPT-6.1 Terra better than Claude Sonnet 5.5?

In my Polish-language test, they tied: 65 to 64.5 out of 72 points. Terra scored 50/51 on office tasks, while Sonnet scored 48.5/51. Terra costs half as much per token.

Which model has lower API costs?

Terra costs $1 / $5 per million tokens, compared with $2 / $10 for Sonnet 5.5. My 12 tasks cost $0.064 on Terra and $0.168 on Sonnet.

Which model writes better in Polish?

Terra lost most of its points on the logic puzzle and code comments. Sonnet lost points on translation, the word limit and an unnecessary proofreading change. For client-facing writing, Terra needed fewer corrections.

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

Sources and test date

• OpenAI: Introducing GPT-6.1 Terra - availability and pricing, checked October 26, 2026
• OpenAI API: GPT-6.1 Terra model - Standard and Batch pricing, checked October 26, 2026
• OpenRouter: API model list - Claude Sonnet 5.5: $2 / $10, Batch $1 / $5, 1-million-token context, checked October 26, 2026
• OpenLEX: tax payment relief - wording of Article 67a § 1, points 2 and 3 of the Polish Tax Ordinance, checked October 26, 2026
• My test: OpenRouter, medium reasoning, October 26, 2026, approximately 20:05-21:45; judges GPT-6.1 Sol and Gemini 4 Argon

Kinetyra