GPT-6.1 Terra vs. GPT-6 Astra: Speed and Cost Come With Accuracy Trade-Offs

Oct 30, 2026 951 views
Glowing cyan wireframe mountains flank a grid valley leading toward a bright sun beneath a purple sky.

GPT-6.1 Terra’s appeal is easy to quantify: across three high-effort tests, it completed the work in 258.8 seconds for an estimated $0.61. GPT-6 Astra took 709.2 seconds and cost an estimated $7.88. The harder question is what that discount buys. In this set of synthetic tasks, Terra scored 121 out of 150 points; Astra scored 148.

That gap wasn’t consistent across every task. The models tied on incident log triage, while Astra led on SQL reasoning and spreadsheet validation. My read is that Terra may suit some workloads where cost and turnaround matter and outputs can be checked. Its mistakes on numeric and cross-record tasks make unsupervised use harder to justify from these results alone.

Price and speed set the stakes

OpenAI released GPT-6.1 Terra on October 26, presenting it as a lower-cost model for high-volume, everyday work. It sits in the middle of the GPT-6 family, below the GPT-6 Astra flagship. OpenAI lists Terra at $1 per million input tokens and $5 per million output tokens, compared with Astra’s $10 and $50.

OpenAI also says Terra completes tasks more than three times faster than Astra on its Agents’ Last Exam benchmark. Those computer-use results aren’t directly comparable to the tests here. To examine the trade-offs on other work, I ran the two models against the same SQL, data validation, and incident triage tasks.

Other early evaluations point to strengths and limits, though their results come from different methods. Plotly reported that Terra found three of five planted patterns in an analytics test and answered 40 of 43 questions correctly. Artificial Analysis found Terra guessed on 38% of questions it couldn’t answer, compared with 88% for GPT-5.6 Terra in the same benchmark. DualEntry reported 78.2% accuracy on its accounting benchmark.

Three tests, one run apiece

I used the final version of each test once per model through OpenAI’s API, setting reasoning effort to high. I didn’t impose an output-token cap, although both models have built-in output limits. Each test had a grading key I built and checked after developing three versions. These results describe single runs on synthetic tasks; they don’t measure repeatability or predict performance across all workloads. Setting the same effort level also doesn’t give both models an identical reasoning-token budget.

The tests covered different kinds of work. The SQL task asked 25 questions about a fabricated database of 500 software builds, including deleted records, retries, multiple currencies, and time zones. Two questions had no answer in the data. The spreadsheet task supplied 4,950 fictional customer records and 19 validation rules. Forty records broke a rule, while 10 records that looked suspicious were actually valid. The incident test presented about 7,000 lines of logs from an outage spanning more than 20 services, with misleading error messages, clock drift, and unrelated changes planted as distractions.

Both models had earned the same scores on earlier easy and medium versions. I made later versions harder after those ties, and continued until GPT-5 Mini scored low. That selection process means the final tasks aren’t a representative sample of ordinary work, and Mini’s results serve only as a comparison point.

For the measurements below, output tokens include reasoning tokens; the parenthetical reasoning count is a subset. Cost estimates use reported tokens at standard uncached input and output rates. They exclude tool charges and billing adjustments, so they aren’t invoice totals.

SQL exposes Terra’s weakest result

Astra scored 48 out of 50 on SQL; Terra scored 26. Each question was worth two points, with credit only for an exact answer. For the two unanswerable questions, the model needed to say the data didn’t support an answer. A blank response earned no points.

Terra processed the 500 rows and handled the deleted builds, retries, and three currencies. On six questions that required aggregating costs or durations across all the rows, though, it supplied SQL queries and marked the totals as not calculated. That response avoided inventing figures, but left the aggregation itself to the user.

Terra also gave six wrong answers, including four counts that were off by one or two. It presented those figures without a signal that they might be uncertain. In a report or operational workflow, small errors can be difficult to catch when they arrive with the same confidence as correct results.

Astra answered more questions correctly but still missed a total. The correct run time for 151 builds in one region was 31,428 seconds; Astra returned 31,022, a difference of 406 seconds, or about 1.3%. The result doesn’t reveal whether it skipped a row, misread a value, or made an arithmetic error. Astra used 27,970 reasoning tokens and took 319.7 seconds.

  • Terra: 133.1 seconds; 21,060 input tokens; 11,554 output tokens, including 10,707 reasoning tokens; estimated cost $0.079.
  • Astra: 319.7 seconds; 21,060 input tokens; 28,806 output tokens, including 27,970 reasoning tokens; estimated cost $1.65.

Validation favors Astra; triage ends in a tie

On data validation, Astra scored a perfect 50 and Terra scored 45. The grading awarded points for finding each of the 40 invalid records and deducted points for wrongly flagging any of the 10 valid ones. Terra caught 35 invalid records and incorrectly flagged none of the valid records: 87.5% recall and 100% precision in this run. Its score is a result on this particular rubric, not a general accuracy rate across all 4,950 records.

Three of Terra’s five misses depended on comparing one record with others, such as spotting a customer who reused someone else’s phone number. The results identify what it missed, but don’t establish why. Astra caught all 40 invalid records, taking 145.0 seconds and an estimated $3.01. Terra took 31.4 seconds and cost an estimated $0.25. It used fewer reasoning tokens, but this single run doesn’t show that token use caused its misses.

  • Terra: 31.4 seconds; 234,045 input tokens; 2,336 output tokens, including 1,841 reasoning tokens; estimated cost $0.25.
  • Astra: 145.0 seconds; 234,045 input tokens; 13,369 output tokens, including 12,831 reasoning tokens; estimated cost $3.01.

Incident triage produced the clearest case for Terra in these tests: both models scored 50. They traced the outage to a settings change in the ledger service, identified three servers with incorrect clocks, and avoided blaming false leads. Those distractions included a rolled-back fraud service update, a payments setting changed shortly before the real cause, a database switchover, and an unrelated memory issue in the search service.

Terra completed the triage in 94.3 seconds for an estimated $0.29. Astra took 244.5 seconds and cost an estimated $3.22. Astra generated 4,628 more reasoning tokens, most of the 6,650-token difference in total output, but both received the same score. These runs can’t tell us whether the extra reasoning helped Astra on the other tasks.

  • Terra: 94.3 seconds; 251,651 input tokens; 7,508 output tokens, including 5,114 reasoning tokens; estimated cost $0.29.
  • Astra: 244.5 seconds; 251,651 input tokens; 14,158 output tokens, including 9,742 reasoning tokens; estimated cost $3.22.

The trade-off depends on the work

TestTerra (out of 50)Astra (out of 50)GPT-5 Mini (out of 50)
SQL reasoning26486
Data validation455010
Incident log triage505012
Total (of 150)12114828
Estimated token cost$0.61$7.88—
Total time258.8 s709.2 s—

Across these runs, Astra took about 2.74 times as long as Terra. Terra’s estimated cost was about 7.8% of Astra’s, while its score was 27 points lower. The results demonstrate a cost, time, and score trade-off on these tests. They don’t show that OpenAI deliberately reduced Terra’s reasoning to reach its price or speed.

For teams considering the models, the practical question is how much checking a task can support. Terra’s incident triage matched Astra here at a fraction of the estimated cost. Its SQL answers and missed records make it a riskier choice for unchecked aggregation or exhaustive validation in this test. Astra also produced a wrong SQL total, so the higher score doesn’t remove the need to verify important outputs. These results suggest that model choice should follow the task’s error tolerance, with measurement on the team’s own data before relying on either system.

The post GPT-6.1 Terra vs. GPT-6 Astra: One sacrifices accuracy for lower cost and higher speed at high effort appeared first on The New Stack.

Source: Jessica Wachtel · thenewstack.io

Comments

Sign in to comment.
No comments yet. Be the first to comment.

Related Articles

GPT-6.1 Terra vs. GPT-6 Astra: One sacrifices accuracy fo...