GPT-6.1 Terra vs Claude Opus 5.5: my English test and real costs
GPT-6.1 Terra came within half a point of Claude Opus 5.5 on my 12 English tasks and cost about 7.6 times less. The details, task by task.

👁 117 reads
I gave GPT-6.1 Terra and Claude Opus 5.5 the same 12 tasks in British English. Opus scored 71.5 out of 72, Terra 71. Opus cost $0.384 for the run, Terra $0.050, so roughly 7.6 times less for half a point. In English, at least, the price gap is far bigger than the quality gap.
- Score: Opus 5.5 71.5/72 (1st of 9 models), Terra 71/72 (tied 2nd with GPT-6.1 Sol, GPT-6 Astra and DeepSeek V4.1 Flash).
- Cost of 12 tasks: $0.384 vs $0.050.
- Output tokens: 9,285 vs 2,704. Opus writes about 3.4 times more.
- API price per 1M tokens: Opus $4 / $20, Terra $1 / $5.
How I tested
I tested Terra on 26 October 2026 through OpenRouter on OpenAI’s Standard endpoint, reasoning effort medium, max 16,000 tokens, one attempt per task. Opus 5.5 answered the same tasks on 24 October with the same settings on Anthropic’s endpoint, and I re-scored all answers in one batch. The tasks were written in British English with UK details: a PCN appeal to Nottingham City Council, idioms, a Cardiff transport news summary, translation traps, VAT sums, date formats, an NHS number validator, a logic puzzle, JSON extraction, UK facts, about 30,000 tokens of meeting minutes and proofreading. Code, maths, JSON and facts were checked automatically; writing was scored blind by GPT-6.1 Sol and Gemini 4 Argon. Terra’s answers were scored by Gemini only, Opus got the average of both judges.
| Metric | GPT-6.1 Terra | Claude Opus 5.5 |
|---|---|---|
| Score / 72 | 71.0 | 71.5 |
| Cost, 12 tasks (USD) | 0.050 | 0.384 |
| Output tokens | 2 704 | 9 285 |
| Total time (s) | 51 | 99 |
| Median tok/s | 149 | 172 |
| Task | GPT-6.1 Terra | GPT-6.1 Sol | Claude Opus 5.5 |
|---|---|---|---|
| Formal letter | 6.0 | 6.0 | 5.5 |
| Idioms and wordplay | 6.0 | 6.0 | 6.0 |
| News summary | 6.0 | 6.0 | 6.0 |
| Translation with traps | 6.0 | 6.0 | 6.0 |
| VAT and currency maths | 6.0 | 6.0 | 6.0 |
| Date and number formats | 6.0 | 6.0 | 6.0 |
| Code: NHS number check | 5.0 | 5.0 | 6.0 |
| Logic puzzle | 6.0 | 6.0 | 6.0 |
| JSON extraction | 6.0 | 6.0 | 6.0 |
| Local facts | 6.0 | 6.0 | 6.0 |
| Long context (~30k tokens) | 6.0 | 6.0 | 6.0 |
| Proofreading | 6.0 | 6.0 | 6.0 |
Where Opus won and where Terra won
Code comments (Opus +1). Both NHS number validators passed every hidden test: Terra accepted 5 of 5 valid numbers and rejected 6 of 6 invalid ones. The judge docked Terra a point because its docstring was a single sentence and the modulus 11 steps had no explanation. Opus documented the weights and the “10 means invalid” rule properly.
Formal letter (Terra +0.5). Opus’s PCN appeal lost half a point: one judge flagged an unsupported claim that the council had suffered no loss, which the brief never mentioned. Terra stuck to the facts given (permit R-77310 slipped onto the dashboard, worn bay markings, request to hold the discounted rate) and scored 6/6.
Same score, different style. Terra’s PCN letter ran to 156 words, Opus’s to 209. In the 30,000 tokens of residents’ meeting minutes both found the current bike store code (9068), the name Hargreaves and the right penalty of £375 per day. Terra gave the three answers in 8 words; Opus spent 100 words walking through meeting numbers and pointing out that code 5530 belongs to the pushchair store. If you need the reasoning shown, Opus provides it. If you only need the answer, Terra charges you less for it.
Everything else: a draw. Both got the VAT sums, UK date formats, logic puzzle, JSON fields, UK facts and all three needles in the long minutes right. The 12-task total differs by half a point, which with one attempt per task is a tie.
The gap is wider in other languages. On the same test in German, Opus scored 71 and Terra 65; in Turkish 72 vs 61. English is where Terra comes closest.
OpenAI’s 9x claim vs my bill
According to OpenAI, on Agents’ Last Exam at max effort Terra is over 3x more token-efficient than Claude Opus 5.5 at a quarter of its price per token, for roughly 9x lower cost per task. My test is different (medium effort, no tools, short tasks), but the direction matches: Terra wrote 3.4 times fewer tokens and the bill was 7.6 times smaller.
On raw ability, the independent Artificial Analysis Intelligence Index (read 26 October 2026) still puts Opus 5.5 well ahead: 57.6 against 48.1 for Terra on max effort. Expect the gap to show on harder, longer reasoning than my office-style tasks.
When to use which
- High-volume API work in English (invoice extraction, form filling, support replies): Terra. 1,000 requests with 2,000 input and 500 output tokens cost $4.50 on Terra and $18.00 on Opus 5.5.
- Long, careful documents where one wrong claim is costly: Opus, or at least a human check. Terra’s answers are shorter and thinner on explanation, as the code comments showed.
- Hard coding: neither choice is obvious from my test. OpenAI itself recommends GPT-6.1 Sol or GPT-6 Astra over Terra for the hardest code. See also my GPT-6 Astra vs Claude Opus 5.5 tests.
- Inside ChatGPT: Terra runs only in Work and Codex (Plus, Pro, Enterprise, Edu), not in Chat.
Full prices, availability and the nine-model ranking are in the GPT-6.1 Terra hub.
Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →Sources
- OpenAI: Introducing GPT-6.1 Terra - read 26 October 2026
- OpenAI API: GPT-6.1 Terra model page - read 26 October 2026
- OpenRouter: Claude Opus 5.5 pricing - read 26 October 2026
- Artificial Analysis: GPT-6.1 Terra - read 26 October 2026
More on this from QUANTSLANT
news·GTA series history: every main game from 1997 to GTA 6 and what it changedEight main games in 29 years: dates, cities and the one big idea each GTA added.

