OpenAI introduced GPT-6.1 Terra on October 26 as a budget-friendly alternative within its GPT-6 lineup, priced at $1 per million input tokens and $5 per million output tokens—one-tenth of GPT-6 Astra's $10 and $50 rates. The company positions Terra for high-volume routine tasks, claiming it completes work more than three times faster than Astra on its Agents' Last Exam benchmark for computer use.

Early independent benchmarks painted a mixed picture. Plotly's analytics test showed Terra identifying three of five planted patterns and answering 40 of 43 questions overall. Artificial Analysis found that Terra guessed on 38% of unanswerable questions instead of declining to answer, a significant improvement from GPT-5.6 Terra's 88% guessing rate. DualEntry's accounting benchmark measured Terra's accuracy at 78.2%.

To examine the speed, cost, and accuracy trade-offs directly, three synthetic tasks were run against both models with reasoning effort set to high. The tests were designed to be challenging after earlier iterations showed both models returning identical scores on easier versions. These single-run results on synthetic workloads do not represent performance across all use cases or establish repeatability.

The tests

Each test was developed through multiple iterations and graded against verified answer keys. Reasoning tokens are included in output totals shown below; the parenthetical reasoning count represents a subset, not an additional charge. Cost estimates are calculated from reported tokens at standard uncached rates and exclude tool charges or billing adjustments.

  • SQL reasoning: 25 questions against a fake database of 500 software builds, including deleted records, retries, three currencies, and timezone-dependent questions. Two questions cannot be answered from the data.
  • Data validation: A spreadsheet of 4,950 fake customer records checked against 19 rules, with 40 records violating rules and 10 appearing incorrect but actually valid.
  • Incident log triage: Approximately 7,000 lines of logs from a simulated outage across 20+ services, with planted errors including clock drift, misattributed error messages, and unrelated alerts preceding the actual cause.

SQL reasoning

This test produced the largest performance gap. Astra scored 48 out of 50 points while Terra scored 26 out of 50. Each of the 25 questions carried two points; answers had to be exact to earn full credit, and unanswered questions earned zero.

Terra processed all 500 rows correctly and handled soft deletes, retries, and currency conversions. Its single numerical error involved calculating total runtime for 151 builds in one region. On six questions requiring aggregation across all 500 rows, Terra declined to calculate answers and instead provided SQL queries for manual execution—marking them as not calculated. While this avoided fabricated numbers, it shifted the computational burden to the user.

Terra answered six questions incorrectly, with four being count errors off by one or two. The model expressed the same confidence in wrong answers as correct ones, providing no signal for verification. Astra's answer to the same region runtime question was 31,022 seconds against the correct 31,428—approximately 1.3% below the target—but still required verification.

  • Terra: 133.1 s, 21,060 input tokens, 11,554 output tokens (10,707 reasoning), $0.079
  • Astra: 319.7 s, 21,060 input tokens, 28,806 output tokens (27,970 reasoning), $1.65

Data validation

Astra achieved a perfect 50-point score while Terra scored 45. Points came from flagging 40 invalid records (one point per correctly identified record with the rule name) and correctly leaving 10 valid records unflagged (one point per correct record).

Terra warned before responding that it could not reliably check all 4,950 records and might miss some violations. It flagged 35 of 40 invalid records while correctly leaving all valid records untouched—87.5% recall and 100% precision in this run. The five missed records only violated rules when compared against other records in the file, such as reused phone numbers. Astra caught all 40 violations but required more than four times the execution time and approximately 12 times the cost.

  • Terra: 31.4 s, 234,045 input tokens, 2,336 output tokens (1,841 reasoning), $0.25
  • Astra: 145.0 s, 234,045 input tokens, 13,369 output tokens (12,831 reasoning), $3.01

Incident log triage

Both models scored 50 points on this test, which required tracing the outage to a settings change in the ledger service, identifying three servers with incorrect clocks, and avoiding false leads planted in error messages. False leads included a rolled-back fraud service update, a payments setting change, a database switchover, and an unrelated memory issue in the search service.

Both models traced the outage correctly and spotted the clock drift issues without accepting error messages at face value. Astra cost approximately 11 times more and took 2.6 times longer to complete the analysis. The difference in output tokens was primarily reasoning tokens: 4,628 of the 6,650-token difference.

  • Terra: 94.3 s, 251,651 input tokens, 7,508 output tokens (5,114 reasoning), $0.29
  • Astra: 244.5 s, 251,651 input tokens, 14,158 output tokens (9,742 reasoning), $3.22

Results

Across the three final high-effort runs, Terra generated fewer reasoning tokens and completed tasks faster than Astra. Combined execution time totaled 258.8 seconds for Terra versus 709.2 seconds for Astra—a 2.74x difference. At standard uncached rates, estimated costs were $0.61 for Terra and $7.88 for Astra, placing Terra at approximately 7.8% of Astra's cost.

Scoring totals were 121 out of 150 for Terra and 148 out of 150 for Astra. The results demonstrate a measurable trade-off in cost, speed, and accuracy but do not indicate whether OpenAI deliberately constrained reasoning to achieve lower costs.

Terra matched Astra's performance on earlier easy and medium test versions, though this does not establish its reliability across all tasks at those difficulty levels. For unchecked numerical aggregation or exhaustive data validation work, caution is warranted. Terra may serve as an acceptable lower-cost option for certain workloads—such as the incident triage scenario tested here—provided results are verified. Astra's incorrect SQL total serves as a reminder that even the flagship model requires verification.