On our two task sets, GPT-6.1 Terra’s low, medium and high effort settings stayed within 7% of each other in cost per round; xhigh cost 1.07 times medium on the easy kit and 1.49 times on the hard kit. Every level got every hard-kit run fully right, and the only miss was one low-effort drawing run on the easy kit; Codex’s default for GPT-6.1 Terra is low. The dollar figures show what the same tokens would cost at API Standard prices: we used a ChatGPT sign-in and paid nothing per token.
- How was GPT-6.1 Terra tested at different reasoning efforts in Codex?
- How did GPT-6.1 Terra perform on the easy kit at low, medium, high and xhigh effort?
- How did GPT-6.1 Terra perform on the hard kit at low, medium, high and xhigh effort?
- Why did GPT-6.1 Terra’s cost barely change between low, medium and high effort?
- What did we not measure?
- FAQs
How was GPT-6.1 Terra tested at different reasoning efforts in Codex?
We ran gpt-6.1-terra at four reasoning effort settings across two task sets for three rounds on a Linux server, using Codex CLI 0.159.0 in non-interactive mode through codex exec on October 28, 2026. Each run started in a fresh folder, with one run at a time. A round contained one run of each task, and the round figures are medians of the three rounds.
The easy kit reused the tasks from our earlier GPT-6.1 Terra cost test:
- A Python code fix checked by 8 tests.
- 10 questions about our own mounting-plate technical drawing, attached to the prompt.
- A 15-field supplier form completed from an email.
We built the hard kit on October 28, 2026, with every right answer computed by our own script:
- A Python package containing 7 bugs, with 16 visible tests and 8 hidden tests. Initially, 4 visible tests failed. The prompt told the model that hidden tests existed, but the model never saw the tests themselves. A run counted as fully right only when all visible and hidden tests passed and the visible tests remained unchanged.
- Our own bracket drawing in two views, including a bolt circle, counterbores, a slot, a title block and a revision table. It had 12 questions, including 4 multi-step calculations.
- A 5-message supplier email thread with later corrections, plus a CSV price list. Its 20 form fields included corrections and calculations.
Codex does not offer none effort, and we did not test max or ultra.
The US dollar estimates use OpenAI’s GPT-6.1 Terra API Standard prices, read on October 28, 2026. Codex showed no cache-write tokens, so we priced uncached input at the 30-minute cache-write rate, 1.25 times the input rate.
For every 1 million tokens, those US dollar rates are:
- Input: $1.00.
- Cached input: $0.05.
- Cache writes with 30-minute retention: $1.25.
- Output: $5.00.
How did GPT-6.1 Terra perform on the easy kit at low, medium, high and xhigh effort?
GPT-6.1 Terra got every easy-kit run fully right at medium, high and xhigh; low got 8 of 9 fully right. Median time per round ranged from 31.5 s at medium to 40.4 s at xhigh.
Easy-kit accuracy, cost and time
| Effort | Runs fully right | One round, median cost | Median time per round | Median output tokens per round |
|---|---|---|---|---|
| low | 8 of 9 | $0.0657 | 32.2 s | 749 |
| medium | 9 of 9 | $0.0630 | 31.5 s | 872 |
| high | 9 of 9 | $0.0668 | 38.9 s | 960 |
| xhigh | 9 of 9 | $0.0677 | 40.4 s | 1,225 |
The one miss at low effort was a drawing run that returned net area 7,799 mm2 and mass 252.7 g instead of the correct 8,999 mm2 and 291.6 g.
How did GPT-6.1 Terra perform on the hard kit at low, medium, high and xhigh effort?
Every effort level got all 9 hard-kit runs fully right, including all 8 hidden code tests. Median time per round was 51.5 s at low, compared with 44.3 s at medium.
Hard-kit accuracy, cost and time
| Effort | Runs fully right | One round, median cost | Median time per round | Median output tokens per round |
|---|---|---|---|---|
| low | 9 of 9 | $0.0733 | 51.5 s | 1,397 |
| medium | 9 of 9 | $0.0689 | 44.3 s | 1,395 |
| high | 9 of 9 | $0.0734 | 48.5 s | 1,642 |
| xhigh | 9 of 9 | $0.1025 | 48.8 s | 1,733 |
Hard-kit median cost per task across three runs was $0.0275 at low and $0.0464 at xhigh for the code fix, $0.0219 and $0.0198 for the drawing, and $0.0235 and $0.0352 for the document.
At medium, the hard-kit medians across three runs were $0.0265 for the code fix, $0.0195 for the drawing and $0.0219 for the document. At high, those task costs were $0.0294, $0.0222 and $0.0219, respectively.
Why did GPT-6.1 Terra’s cost barely change between low, medium and high effort?
Most of the cost came from input, which Codex sends again at every step; output accounted for 6.1% to 8.7% of easy-kit cost and 9.5% to 11.3% of hard-kit cost. Cost moved more with the number of steps than with the effort setting: low, medium and high stayed within 7% of each other per round on both kits, while xhigh cost 1.07 times medium on the easy kit and 1.49 times on the hard kit. Individual runs varied, including one high-effort form run that took extra steps and cost $0.0592, against about $0.02 for the others.
What did we not measure?
The runs did not measure the following:
maxandultrareasoning effort.nonereasoning effort, which Codex does not offer.- Runs through the API itself.
- Tasks outside the easy and hard kits.
FAQs
What is GPT-6.1 Terra’s default reasoning effort in Codex?
Codex lists low as GPT-6.1 Terra’s default reasoning effort.
Does high effort cost much more than medium for GPT-6.1 Terra?
In our runs, low, medium and high stayed within 7% of each other per round on both kits. Easy-kit median cost was $0.0630 at medium and $0.0668 at high; hard-kit median cost was $0.0689 and $0.0734, respectively, at API-equivalent rates.
Was xhigh more accurate than medium for GPT-6.1 Terra on the hard kit?
No. Both medium and xhigh got all 9 hard-kit runs fully right, including all 8 hidden code tests.
Did low effort get every GPT-6.1 Terra run fully right?
No. At low effort, GPT-6.1 Terra got 8 of 9 easy-kit runs and all 9 hard-kit runs fully right. The easy-kit miss was a drawing run with incorrect net-area and mass calculations.
Can I use none effort for GPT-6.1 Terra in Codex?
No. Codex does not offer none for GPT-6.1 Terra, although OpenAI’s API supports it.
For API calls, reasoning.effort accepts these settings:
nonelowmedium, the API default.highxhighmax
Were the GPT-6.1 Terra test runs billed per token?
No. We used a ChatGPT sign-in in Codex and paid nothing per token. The dollar figures represent what the recorded token counts would cost at API Standard prices.
Related on NerdsChalk
Leave a Reply