GPT-6.1 Terra scored 121 against Astra's 148 in three hard tasks

A new OpenAI model priced at one tenth of the top one does not come free. Jessica Wachtel at The New Stack ran three hard tasks against GPT-6.1 Terra. She compared each run with the full GPT-6 Astra, reasoning set to high. Terra scored 121 of 150 and Astra 148 of 150. The runs cost $0.61 for Terra and $7.88 for Astra at listed rates.

OpenAI released Terra on 26 October. The company lists Terra at $1 per million input tokens and $5 per million output. That is one tenth of Astra's $10 and $50. OpenAI also says Terra finishes computer-use tasks three times faster than Astra. That claim comes from its Agents' Last Exam benchmark. The three tests below are not that benchmark.

What each test asked

The SQL test gave each model a fake database of 500 software builds and 25 questions. Some rows were deleted, some were retries, costs were in three currencies, and a few questions crossed time zones. Two of the questions could not be answered from the data.

The data validation test gave each model a spreadsheet of 4,950 fake customer records and 19 rules. Some rules only make sense across the whole file. Forty records broke a rule and ten records looked wrong but were fine.

The incident log test gave each model about 7,000 lines of logs from a fake outage. The logs held three false leads. A rolled-back fraud update. A payments setting that changed seconds before the real cause. An unrelated memory problem in the search service.

Where Terra fell short

On the SQL test, Terra scored 26 and Astra 48. Six questions needed sums or counts across all 500 rows. Terra wrote a SQL query for each and asked the reader to run it. Wachtel calls that the first time she has seen a model hand work back. Terra also got six questions wrong. Four of them were counts off by one or two. Terra gave those answers with the same confidence as its correct ones.

On the data validation test, Terra scored 45 and Astra 50. Terra caught 35 of the 40 invalid records and flagged no valid ones. Astra caught all 40. Terra used 1,841 reasoning tokens and finished in 31 seconds. Astra used 12,831 and took 145 seconds.

On the incident log test, both models scored 50. Both traced the outage to a settings change in the ledger service. Both found the three servers with wrong clocks. Neither blamed the planted false leads. Astra cost about eleven times as much and took 2.6 times as long. The extra output was mostly reasoning tokens.

The pattern across the three

Terra generated fewer reasoning tokens and finished sooner on every test. On the SQL and data validation tests, the lower score and shorter time came together. On the incident log test they did not. The runs show a trade-off between cost, time, and score. They do not show whether OpenAI cut reasoning on purpose, Wachtel writes.

Wachtel rebuilt each test several times after both models kept returning identical scores on easier versions. She picked the hard versions after seeing results, so these are single runs and not a sample of everyday work. GPT-5 Mini scored 28 of 150 as a baseline.

For a team choosing a model by the task, the useful finding is where Terra's errors sat. They were in numerical aggregation and in cross-record rule checks. A buyer paying one tenth of the price can accept that on incident triage. On exhaustive data checks or unchecked sums, a verified answer still needs the slower model.

The New Stack notes that its owner Insight Partners is an investor in OpenAI.

Source

GPT-6.1 Terra vs. GPT-6 Astra: One sacrifices accuracy for lower cost and higher speed at high effort, by Jessica Wachtel, The New Stack.