GPT-6 Luna, Terra, and Sol: A Tie on Checkable Tasks

On November 4, 2026, a benchmark test revealed that GPT-6 Luna, GPT-6.1 Terra, and GPT-6.1 Sol each answered all 25 checkable tasks correctly, achieving a perfect score of 100% [1]. This result means that these specific tasks can no longer differentiate between the models based on accuracy alone. However, further examination of their performances on a more challenging set of tasks highlights differences that could inform user decisions.

Performance on Standard Tasks

The recent testing conducted on November 4, 2026, evaluated three models across a set of 25 checkable tasks, with results meticulously gathered to ensure clarity and reliability. Here is each model's result from the scoreboard [1]:

All three models, GPT-6 Luna, GPT-6.1 Terra, and GPT-6.1 Sol, achieved perfect scores on the 25 tasks, meaning they all passed without distinction. This result indicates that standard benchmarks currently do not provide a clear advantage among these models based solely on this particular task set.

However, it's important to note that while they excelled on these 25 tasks, this small set may not adequately reflect their performance in more complex queries. For instance, how well they tackle nuanced questions or intricate problem-solving scenarios remains untested by this benchmark.

Additionally, the cost per run varies significantly, with GPT-6 Luna being the most economical option. Luna (1.28 seconds) and Terra (1.32) differ by less than 0.1 seconds, which is noise; Sol was slower at 1.80 [1].

Deeper Insights from Harder Tasks

While the initial task set provided no differentiation among the models, a follow-up assessment with a more challenging group of 7 tasks revealed distinct insights into their capabilities. Here are the results of this second set, from the same scoreboard [1]:

In this analysis, GPT-6.1 Terra and GPT-6.1 Sol both achieved perfect scores, while GPT-6 Luna scored 86% on a more challenging set of tasks that included trap questions. Notably, Luna previously scored between 86% and 95% in different runs of the same tasks, highlighting the variability that can occur under different testing conditions. Its score on the trap questions, the ones with a tempting wrong answer, was 87%, against 100% for Terra and Sol [1].

Cost Efficiency and Speed

Cost and speed are critical considerations when selecting a model for tasks. Our own data from a test conducted on 4 November 2026 gives these costs for 100 tasks of the harder set [1]:

On the harder set the median response times were 1.58 seconds for Terra, 2.12 for Luna and 2.23 for Sol; on the standard set Luna (1.28 seconds) and Terra (1.32) were within 0.04 seconds of each other and Sol took 1.80 [1]. Seconds are wall time from one PC, network included.

It’s important to note that while all three models scored 100% on the standard 25 tasks, the results from the harder 7-task set provide a clearer distinction, where GPT-6 Luna scored 86% compared to 100% for both Terra and Sol. This variability suggests that performance can fluctuate across different task sets and conditions, highlighting the need for multiple runs when assessing model efficacy.

Questions readers ask

Why are the 25 tasks not sufficient to differentiate the models?

On 25 tasks one missed question is 4 points, so models within a few points of each other are tied, and three models at 100% cannot be ranked [1].

How do the models compare on cost efficiency?

GPT-6 Luna was the cheapest model on both sets: $0.0005 per run on the standard set and $0.0076 per 100 tasks on the harder set, against $0.0400 for Terra and $0.0648 for Sol. It also scored lower on the harder set (86% against 100%), so the lower price comes with that trade-off [1].

Are there any risks in relying solely on task scores?

Yes, while scores provide a snapshot of performance, they do not account for the variability in real-world applications. It’s advisable to consider multiple runs and a range of tasks to get a comprehensive picture of each model's capabilities.

How the numbers work

These figures were computed for this article, not estimated by its writer. Inputs come from the numbered sources or are assumptions, labelled as such; change an assumption and the results change with it.

Sources

1. Our own data, 4 Nov 2026: Model bench: 25 checkable tasks, scored by code