Artificial Analysis's independent evaluation shows that GPT-6.1 Terra scores higher on the Intelligence Index at a lower task cost, with low and medium reasoning on the cost-efficiency frontier; low and xhigh are on the frontier in the Coding Agent Index. This conclusion applies to these benchmarks and the cost methodology used on this page; it does not directly establish value for money in other workflows.
Tasks this can help assess: Compare GPT-6.1 Terra with GPT-5.6 Terra on Artificial Analysis indices, reasoning levels, and estimated task costs; understand changes in terminal coding and factuality benchmarks.
Tasks this should not be extrapolated to: Specialized tasks not included in the indices, success rates in real applications, actual end-user bills, or score comparisons across different evaluation frameworks.
Applicable model versions: GPT-6.1 Terra; the article compares it with GPT-5.6 Terra, GPT-6.1 Sol, GPT-6 Luna, and other models.
Test environment or client: Artificial Analysis Intelligence Index v4.3.2, Coding Agent Index v2.0, and the individual benchmarks listed in the article; the runtime environment, prompts, tool configuration, and full evaluation logs are not disclosed.
Reasoning levels and parameters: The charts compare settings including low, medium, high, xhigh, max, and non-reasoning; temperature, sampling parameters, and the specific configuration for each level are not specified.
Intelligence Index v4.3.2: The page says the index combines 10 evaluations: AA-Briefcase v1.1, GDPVal-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity's Last Exam, GDP.pdf, CritPt, AA-Omniscience, and AA-LCR v1.1. The chart notes that no public API is available.
Coding Agent Index v2.0: The chart says it consists of three benchmarks: FrontierCode v1.1, Terminal-Bench 4.0, and SWE-Atlas-QnA-AA. The cost chart uses the average per-task API cost, priced by token, on the x-axis; the index and cost charts show scores and costs for different models or reasoning levels.
AA-Omniscience: The page describes this as Artificial Analysis's private knowledge and hallucination benchmark. The chart says the index rewards correct answers and penalizes hallucinations, without penalizing refusals; answer accuracy and hallucination rate are also reported.
Cost comparison: The article gives API prices per million tokens and compares index performance using weighted average task costs. It does not disclose the input/output token distribution for tasks, the weighting formula, evaluation sample size, or number of repetitions.
Terminal-Bench 4.0: The chart notes that it was developed by the Laude Institute and open-source contributors, and independently evaluated by Artificial Analysis; the page does not provide the full runtime configuration.
Intelligence Index and task cost: GPT-6.1 Terra max scores 48, 6 points higher than GPT-5.6 Terra max; their task costs are approximately $0.40 and $1.40, respectively. Low and medium are on the chart's cost-efficiency frontier; the article says lower reasoning levels score 7–13 points higher than GPT-5.6 Terra. The max score is 4 points below GPT-6.1 Sol and 10 points above GPT-6 Luna.
Coding Agent Index: GPT-6.1 Terra scores 48, below GPT-6.1 Sol at 54 and above GPT-6 Luna at 31. The chart shows low and xhigh on the cost-efficiency frontier; the other Terra reasoning levels are not on the frontier.
AA-Omniscience: In the chart, GPT-6.1 Terra max has an index score of 30, answer accuracy of 49%, and hallucination rate of 38%. The article says that compared with GPT-5.6 Terra, the hallucination rate fell by 50 percentage points from 88%, while accuracy rose by 2 percentage points.
Terminal-Bench 4.0: The chart shows GPT-6.1 Terra max at 55.6%, versus 35.4% for GPT-5.6 Terra max, an increase of about 20.2 percentage points; GPT-6.1 Sol max scores 56.1%.
API pricing: GPT-6.1 Terra input/output prices are $1/$5 per million tokens, compared with $2/$12 for GPT-5.6 Terra; cached writes cost $1.25, and cached input costs $0.05, compared with $0.20 for GPT-5.6 Terra.
| Metric | GPT-6.1 Terra | Comparison or basis |
|---|---|---|
| Intelligence Index v4.3.2 (max) | 48 | 6 points higher than GPT-5.6 Terra max; GPT-6.1 Sol is about 52, GPT-6 Luna about 38 |
| Intelligence Index task cost (max) | $0.40 | GPT-5.6 Terra max is $1.40; estimated by Artificial Analysis per task |
| Coding Agent Index v2.0 | 48 | GPT-6.1 Sol 54; GPT-6 Luna 31 |
| AA-Omniscience (max) | Index 30; accuracy 49%; hallucination rate 38% | GPT-5.6 Terra hallucination rate 88%; article reports an accuracy difference of +2 percentage points |
| Terminal-Bench 4.0 (max) | 55.6% | GPT-5.6 Terra 35.4%; GPT-6.1 Sol 56.1% |
| Input / output API prices | $1 / $5 per million tokens | GPT-5.6 Terra is $2 / $12 |
| Cached write / cached input prices | $1.25 / $0.05 per million tokens | Cached input comparison price is $0.20; the article says the cached write price was halved |
The index scores, costs, and individual benchmark scores above come from the article's body and charts. Some values were read from charts; the page does not provide raw per-question results, confidence intervals, or sample sizes. Costs are benchmark task estimates from Artificial Analysis and should not be treated as actual prices paid for any user scenario.
This page provides benchmark results from Artificial Analysis's own evaluation and task-based cost estimates, supporting comparisons among the models, indices, and reasoning levels shown on the page.
The cost-efficiency frontier for each of the two indices is based on the scores and costs in its respective chart: low and medium for the Intelligence Index, and low and xhigh for the Coding Agent Index. The frontier finding should not be transferred across indices.
AA-Omniscience is a private benchmark; the page does not disclose its questions, sample size, or scoring details. The indices and cost calculations also lack full reproduction configurations. Although Terminal-Bench is a public benchmark, the article still does not disclose the runtime environment or invocation settings.
The results reflect a snapshot of the page dated 2026-10-26. They do not represent performance across all tasks; prices, model versions, and dynamic indices may change later.
On 2026-10-27, the specified original page was opened through Tabbit and its date, body, five embedded charts, and chart annotations were checked. The charts show results and some cost data for Intelligence Index v4.3.2, Coding Agent Index v2.0, AA-Omniscience, and Terminal-Bench 4.0. The API, private evaluation questions, full evaluation configuration, sample sizes, and cost-calculation details are not disclosed, so this article transcribes only the results visible on the page; it does not estimate missing data or conduct an independent retest.
GPT-6.1 Terra
