All news
Product

Spark AI at the Medium level is faster and more thorough, with GPT-6.1 Terra

On the day OpenAI released GPT-6.1 Terra, we tested it on real questions from the ornamental horticulture trade. It answered all of them correctly, and faster. That is why the Medium level in Spark AI now runs on Terra, and Medium is the default.

The heading Medium now thinks with Terra, above the Spark AI menu with Fast, Medium and High, where Medium is selected with the label GPT-6.1 Terra

The Medium level in Spark AI now runs on GPT-6.1 Terra, OpenAI's newest model. You will notice answers that arrive faster and that were all correct in our test. Medium is also the new default. You don't need to do anything.

Tested on the same day

OpenAI released GPT-6.1 Terra on 26 October. That same day we tested it on our own test set: 28 real questions from the ornamental horticulture trade, about revenue, stock, purchasing, auction prices and who is allowed to see which figures. We check every answer with our own calculation on the same data. No AI grades itself.

We put Terra next to the models Spark AI already used, at every reasoning level, and had each combination run the full set twice.

What changes

  • Medium is faster. Half of all answers arrive within 19 seconds, down from 27 seconds. The slowest answers arrive sooner too.
  • Medium is more thorough. In two test rounds, Terra answered all 56 questions correctly.
  • Medium is the default. A new conversation starts on Medium. If you only want to look up a number quickly, choose Fast yourself.
  • Fast and High stay as they were.
Two pairs of bars: half of all answers arrive within 19 seconds instead of 27, and 95 percent within 44 seconds instead of 55

With Terra, Medium does use more AI credit per question than before, but still about half of what High uses.

More thinking is not always better

We also tested what happens when you let a model think longer. It did not produce better answers. At the highest level, the computation grew almost ninefold, an answer could take more than three minutes, and the score went down.

Five bars for the reasoning levels of GPT-6.1 Terra. Computation rises from 0.6 times at low to 8.9 times at max, while medium and xhigh answer all 56 questions correctly and max gets 53 of 56 right

That is why we choose the level at which a model answers best, not the most expensive one.

Keeping up without doing it yourself

New AI models are appearing at a rapid pace. Keeping track of which one is better, faster or cheaper, and whether that also holds for questions about your own data, is a job in itself. We do it for you: we test every new model on questions from the day-to-day work of growers and trading companies, and only put it to use once it scores better on those questions.

It is part of your subscription. Packages start at €400 per month, excluding VAT, with AI credit included and no implementation costs.

Read more