GPT-6.1 Terra focuses on low cost, at the price of some errors

OpenAI positions GPT-6.1 Terra as a fast and inexpensive model. Tests published by The New Stack show a clear advantage in cost and time, but also weaknesses in data aggregation.

OpenAI launched GPT-6.1 Terra on October 26, presenting it as an inexpensive model for high-volume everyday tasks. In a comparative test published by The New Stack on October 30, 2026, Terra finished faster and at a lower cost than GPT-6 Astra, but with a lower score on two of the three exercises. For teams automating queries, data checks, or incident sorting, the gap quickly matters.

Le calcul est simple : moins cher, mais pas gratuit en confiance

Terra is priced at 1 dollar per million input tokens and 5 dollars per million output tokens, one-tenth of Astra's rates, billed at 10 dollars and 50 dollars on the same basis. In the three synthetic tests conducted by Jessica Wachtel with reasoning effort set to high, Terra cost a total of 0.61 dollars versus 7.88 dollars for Astra. The total time also favored Terra: 258.8 seconds for Terra, 709.2 seconds for Astra.

However, the picture is not linear. In the SQL reasoning test, Terra scored 26/50 while Astra reached 48/50. In data validation, Terra scored 45/50 and Astra 50/50. Conversely, the two models tied on sorting a false incident, each scoring 50/50. The difference here is not only a matter of price; it concerns reliability when accuracy is required.

Quand le modèle répond, mais laisse le travail à moitié fait

The SQL test best illustrates the dilemma. Terra successfully processed 500 lines of a fictitious database, handled logical deletions, retries, and three currencies, but refused to calculate six answers and returned SQL queries to be executed manually. Jessica Wachtel notes this is the first time she has seen a model “hand work back.” In other words: the AI was not wrong on everything; it sometimes chose not to decide. Useful caution, but not always practical when an actionable result is expected within a minute.

In data validation, Terra identified 35 of 40 erroneous records without false positives. Astra found all 40 but at a higher cost: 145.0 seconds versus 31.4 seconds for Terra, and 13,369 output tokens versus 2,336. In incident sorting, both models followed the correct lead—a parameter change in the ledger service—and avoided false clues, including a canceled fraud change, a database switch, and a false memory issue. Again, the cost gap did not translate into a result gap.

Pour les équipes, le vrai sujet reste le périmètre d’usage

The total of the three tests gives 121/150 for Terra versus 148/150 for Astra. The New Stack reminds that these results come from final versions of synthetic tests, hardened through iterations because models often achieved the same scores. In other words, this is not a general performance measure but a comparison point on specific tasks. This changes the reading of the launch: Terra may suit repetitive workloads where cost matters, but requires human verification whenever numbers must be accurate to the cent or unit.

OpenAI clearly seeks to broaden the use of its models with an intermediate offer between speed and cost. On paper, Terra lowers the entry ticket. In practice, it mainly reminds that a cheaper model remains a model to control. The machine can be fast; the spreadsheet, however, does not always forgive.

By the numbers

  • 1 dollar / 5 dollars per million tokens — GPT-6.1 Terra input/output rates, OpenAI, October 26, 2026.
  • 10 dollars / 50 dollars per million tokens — GPT-6 Astra input/output rates, OpenAI, October 26, 2026.
  • 258.8 seconds — total time for Terra on three tests, The New Stack, October 30, 2026.
  • 709.2 seconds — total time for Astra on three tests, The New Stack, October 30, 2026.
  • 121/150 vs. 148/150 — total scores for Terra and Astra, The New Stack, October 30, 2026.

To read