OpenAI's newly released GPT-6.1 Terra model completed tasks more than three times faster than its flagship GPT-6 Astra but scored substantially lower on numerical reasoning tests, according to a benchmark analysis published October 30 by The New Stack. The company launched Terra on October 26, positioning it as a budget-friendly option for high-volume everyday work at a tenth of Astra's price. Independent testing across three synthetic tasks revealed clear trade-offs between cost, speed, and reliability that could matter for businesses deciding which model to deploy.
In three high-difficulty tests measuring SQL reasoning, data validation, and incident log triage, Terra earned 121 out of 150 points while Astra scored 148 out of 150. The widest gap appeared in the SQL reasoning challenge, where Astra achieved 48 out of 50 points compared with Terra's 26 out of 50. Terra processed all 500 database rows and handled deleted records, retries, and multiple currencies, but it refused to calculate six questions requiring totals across all rows and instead returned SQL queries for the user to run manually. On six other questions, Terra provided confident answers that were wrong—four counts missed the mark by just one or two, with no signal that the numbers were incorrect. The data validation test showed a smaller gap: Terra flagged 35 of 40 invalid records with zero false positives, earning 45 out of 50 points, while Astra caught all 40 and achieved a perfect score. Both models tied at 50 out of 50 on the incident log triage test, correctly identifying a ledger service settings change as the root cause and spotting three servers with incorrect clocks while ignoring multiple false leads planted in the logs.
Measured elapsed time totaled 258.8 seconds for Terra and 709.2 seconds for Astra, meaning Astra required roughly 2.74 times as long overall. At standard uncached rates, the estimated token costs came to $0.61 for Terra and $7.88 for Astra—putting Terra's expense at approximately 7.8% of its flagship counterpart. OpenAI lists Terra at $1 per million input tokens and $5 per million output tokens, one-tenth of Astra's $10 and $50 pricing. Terra generated fewer reasoning tokens in every test: 10,707 versus 27,970 on SQL reasoning, 1,841 versus 12,831 on data validation, and 5,114 versus 9,742 on incident triage. Early third-party benchmarks reflected similar patterns—Plotly reported that Terra identified three of five planted patterns in an analytics test and passed 40 of 43 questions overall, while Artificial Analysis found it guessed on 38% of questions it couldn't answer instead of admitting uncertainty, down from 88% for the earlier GPT-5.6 Terra.
The analysis cautions that Terra may not be suitable for unchecked numerical aggregation or exhaustive data validation, though it could serve as an acceptable lower-cost alternative for some workloads if results are verified. The report notes that these single-run results on synthetic tasks shouldn't be treated as a representative sample of everyday work, since the tests were rebuilt several times and made progressively harder after both models kept returning identical scores on easier versions. Astra's own incorrect SQL total—missing 406 seconds out of a correct 31,428, or about 1.3%—reinforces the need to check its work as well. The measured trade-offs show a cost, time, and score split, but the runs don't reveal whether Terra was overwhelmed by difficulty or why it made specific errors. Organizations weighing the models will need to decide whether Terra's speed and dramatic cost savings justify accepting lower accuracy on complex numerical tasks, or whether Astra's higher reliability is worth paying roughly twelve times more per workload. For teams already building verification processes into their workflows, the cheaper model's performance on tasks like incident triage suggests it won't always be the wrong choice.