GPT-6.1 Terra Reasoning Levels: Low, Medium, High or Max?

👁 118 reads

GPT-6.1 Terra has six reasoning levels: none, low, medium, high, xhigh and max. I tested what each level delivers and what it costs using the same six difficult tasks with one correct answer: Sudoku, a logic puzzle, PIN-code counting, a loan payment, working days and letter counting. I also tested a technical drawing and a letter to a tax authority. With reasoning set to none, Terra solved one of the six tasks. At low, it solved all six. Max cost 7.7 times more than low and did not earn a single extra point.

What OpenAI says and what the benchmarks show

According to the model page, the API offers Terra’s reasoning levels as none, low, medium, high, xhigh and max, with medium as the default. In its GPT-6 model guide, OpenAI recommends starting with medium and changing the level only when results call for it. Higher levels use more reasoning tokens, which are billed at the output-token rate of $5 per million.

The independent Artificial Analysis index scored Terra at every level (checked on 26 October 2026):

Artificial Analysis also reports the cost of one task in its index: $0.095 at medium and $0.40 at max. The index score rises at every level because it includes many very difficult tasks. In my practical tasks, Terra had already reached the ceiling at low.

How I tested the model

Each task had one correct answer, which I calculated in advance with a program, making the scoring automatic and eliminating the need for AI judges. Terra received each task without hints, with a 64,000-token limit, through OpenRouter’s OpenAI endpoint, priced at $1/$5. I ran each level once.

• Logic puzzle: Five people, five cities, five drinks and 13 clues. The questions were who drinks the juice and where Eva is from. I checked that the puzzle had a unique solution by examining every possible arrangement.
• PIN codes: How many six-digit codes have no three identical digits in a row and have a digit sum divisible by 7? The answer is 137,610. The prompt prohibited writing a program.
• Loan payment: A loan of about $76,336 over 25 years at 7.2% annually, with equal payments. The answer is about $549.30.
• Working days: From 1 November to 31 December 2026, excluding holidays. The answer is 41 (11 November, Christmas Eve and 25 December fall on weekdays).
• Sudoku: A puzzle with 21 starting digits. The answer is nine lines of nine digits.
• Letter count: Count the Polish letter a with an ogonek in a paragraph about Agnieszka and a basket of apples. The answer is 11. The original test material was in Polish; this description is a translation, not a retest in English.

Without reasoning (none): 1 out of 6

At none, Terra answers immediately, without hidden deliberation. It got only the working-day count right. Here is how it did on the other tasks:

• Loan payment: The formula was right, but the result was wrong: about $549.40 instead of about $549.30. A difference of about 10 cents per payment adds up to about $31 over the full loan.
• PIN codes: The inclusion-exclusion method was correct, but Terra started with the wrong count of codes whose digit sum was divisible by 7, arriving at “RESULT: 872055.” At low, the same method produced 137,610.
• Sudoku: It got 79 of 81 cells right, but swapped the last two digits in two rows: “369845712” instead of “369845721.”
• Logic puzzle: It decided the clues were contradictory and replied, “JUICE: not determined; EVA: not determined.” The clues have one solution: Celina drinks the juice, and Eva is from Lublin.
• Letter count: “RESULT: 12” instead of 11, in a 10-token response. The original test was in Polish; the quoted output here is translated.

The takeaway: none is suitable for rewriting, summarising and classification, but not for calculations. The same loan calculation cost $0.0005 at none and $0.0015 at low - a tiny amount for a correct result.

From low to max: same score, rising cost

The PIN task makes the difference clear. At low, Terra used 1,232 tokens and took 15 seconds; at max, it used 14,159 tokens and took 157 seconds. Both times, the answer was 137,610. Max is not “more accurate medium”; it spends longer checking the same result.

Technical drawing and letter: more reasoning didn’t help

I tested the same levels on two tasks from my first Terra test.

Flange drawing (13 points): chained dimensions, six holes on a pitch circle, a contradictory thickness (8 mm in the view, 10 mm in the title block) and a mass to calculate. Terra scored 13/13 at every level, including none. The cost rose from $0.0036 at none to $0.0113 at max. Reading the drawing does not require extended reasoning - the model just needs to see it correctly.

Letter to a tax authority: The task was to request a payment plan for overdue tax, where the relevant legal basis in the original Polish test was Article 67a(1)(2) of Poland’s Tax Ordinance. At none, low, high (three attempts) and max, Terra gave the correct provision. At xhigh, it wrote “Article 67a(1)(3),” the provision on cancellation; at medium, it made one mistake in five attempts. A higher level does not prevent this kind of error - it is a matter of randomness, not deliberation. I separately checked how often Terra gets legal provisions wrong, and how to prevent it, in 20 attempts.

How to set the reasoning level

API (OpenRouter or OpenAI Chat Completions):

{
  "model": "openai/gpt-6.1-terra",
  "reasoning": {"effort": "low"},
  "messages": [...]
}

In the OpenAI Responses API, the parameter looks the same: reasoning={"effort": "low"}. Codex: add model_reasoning_effort = "low" to ~/.codex/config.toml. ChatGPT Work: choose the level alongside the model. I compared Terra with GPT-6.1 Sol at different levels in Terra vs Sol vs Luna, and covered the costs in GPT-6.1 Terra pricing.

Which reasoning level should you choose?

• none - text transformations, translation, extracting fields from documents and reading drawings. Not for calculations or puzzles.
• low - high-volume tasks with clear answers: invoices, forms and simple calculations. In my test, it got as many answers right as max for 13% of the cost.
• medium - a sensible starting point when you do not know how difficult a task will be. I used this level to test Terra in my GPT-6.1 Terra guide.
• high, xhigh, max - use these only when medium gets something wrong. They could not improve my results, and max extended the response time to as much as 2.5 minutes.

Frequently asked questions

What is GPT-6.1 Terra’s default reasoning level?

Medium. That is what the OpenAI API model documentation says, and OpenAI recommends starting there.

Does max give better answers than medium?

In the Artificial Analysis index, yes: 48.1 versus 41.3 points. Not in my six practical tasks: low, medium and max all got a perfect score, while max cost five times as much as medium.

How much does reasoning cost?

Reasoning tokens are billed as output tokens: $5 per million. My six tasks cost from $0.0076 at none to $0.1227 at max.

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

Sources and test date

• OpenAI API: GPT-6.1 Terra model - reasoning levels, medium as the default and pricing (checked 26 October 2026)
• OpenAI: Using GPT-6, GPT-6.1 Terra section - recommendation to start with medium (checked 26 October 2026)
• Artificial Analysis: GPT-6.1 Terra - index scores at each level and task costs (checked 26 October 2026)
• My test: 6 tasks at 6 levels, plus a drawing and a letter; OpenRouter, OpenAI endpoint, 26 October 2026, approximately 9:00-9:20 p.m.; total cost $0.39

Kinetyra