Five models ace our 25 tasks; only the trap questions tell them apart

On November 6, 2026, an evaluation revealed that five advanced models, including GPT-6 Luna and GPT-5.4, successfully completed all 25 tasks in a benchmark test, each scoring 100% [1]. This result implies that distinguishing between these models based solely on this test is no longer feasible. This article breaks down their performance, costs, and implications for users choosing among these models.

Detailed Model Performance

The models evaluated were GPT-6 Luna, GPT-6.1 Terra, GPT-5.4, GPT-6.1 Sol, and GPT-5.5. Each model answered all tasks correctly, achieving a perfect score of 100% [1]. However, they differ significantly in terms of cost per run and speed. Below is a comparative table summarizing their performance based on our own test data from November 6, 2026 [1]:

The median response time highlights how quickly each model processed the tasks. Of the five, GPT-5.4 was the fastest, with a median of 0.97 seconds [1]. When considering cost, GPT-6 Luna is the least expensive at $0.000506 per run [1]. However, while this model is economical, its performance on more complex tasks may not be as robust as that of some of the other models, particularly when faced with challenging scenarios, where the slight uptick in cost for a more capable model could yield better results.

Consistency and Reliability

Consistency in performance is a critical factor when evaluating AI models. All five models tested delivered perfect accuracy in the initial evaluation, answering all 25 tasks correctly [1]. Each task was asked three times. Consistency, the share of tasks a model answered the same way in all three runs, was 96% for GPT-5.4, 92% for GPT-6 Luna and GPT-6.1 Sol and GPT-5.5, and 88% for GPT-6.1 Terra [1]: a right answer on one run does not guarantee the same answer on the next.

While the first set of tasks showed no variance in performance, follow-up tests on a second set of more challenging tasks revealed differing capabilities among the models. The second set has 7 tasks built to trip a model: five trap tasks with a tempting wrong answer and two controls without one, each asked three times [1].

Performance Summary on Harder Tasks:

The cheaper and older models fell much further on the traps: GPT-5.4 mini 71%, GPT-4.1 67%, GPT-4.1 mini 57% and GPT-4o mini 29% [1]. GPT-6 Luna missed one trap answer in its runs and stayed close behind the four at 100% while costing far less. With 7 tasks one miss is 14 points before repeats, so read small gaps as ties [1]. No model's score on the standard set moved since the run of 4 November [1].

Cost Considerations

A run is the 25 tasks once, so 1,000 tasks is 40 runs. From our own data of 6 November 2026 [1]:

The cheapest of the five, GPT-6 Luna, is about 6 times cheaper than GPT-6.1 Terra and about 72 times cheaper than GPT-5.5 on these tasks [1].

However, while price is important, it's essential to balance it with performance metrics. For instance, if your project involves more complex tasks where accuracy is paramount, you might consider a model like GPT-6.1 Terra or GPT-6.1 Sol, despite their higher costs. They scored 100% on the trap set where Luna scored 95%: that is the difference the higher price buys on these tasks [1].

Ultimately, the choice of model should reflect both your project's performance requirements and financial constraints, considering not just the base cost but also the expected outcome in terms of task complexity and accuracy.

Questions readers ask

What should I consider when choosing between these models?

Consider the specific tasks you'll be running, the required accuracy, response time, and your budget. For general tasks, GPT-6 Luna is the most cost-effective, while for challenging scenarios, GPT-6.1 Terra or GPT-5.4 may be more appropriate.

Are there any limitations to these models?

While these models achieved perfect scores on the tasks tested, performance may vary with different contexts or more complex questions. Always evaluate them in the context of your specific use case.

How can I test these models for my own needs?

The task list, the scoring code and every answer are published with the scoreboard [1], so the run can be repeated on your own key; better still, add ten tasks from your own work.

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, 6 Nov 2026: Model bench: 25 checkable tasks, scored by code