New OpenAI Models: Do Better Benchmarks Mean Better Work?

Artificial Analysis reports that GPT-6.1 Terra delivers better results at lower cost than its predecessor. In its October 26, 2026 assessment, the model gains six points over the previous Terra release on the Intelligence Index at maximum effort, while average cost per task falls from $1.40 to $0.40. Those results and costs apply to the tasks in that evaluation.

Each new model arrives with a familiar promise: better performance, greater efficiency, a lower bill. Yet that is not what I have experienced with OpenAI’s last few releases. When using GPT-5.6 and GPT-6 versions after GPT-5.5, I have encountered more errors.

I am not presenting a controlled comparison between GPT-6.1 Terra, the subject of that report, and the versions I have used. But there is a gap between the expectation these announcements create and my own experience. Why does switching to a supposedly better model leave me with more problems in the work I do?

A lower price is not enough reason to switch

I would welcome a model that does the same work, just as accurately, in less time and with fewer resources. That would be an improvement. If accuracy declines, though, a lower price alone is harder to count as a gain.

A response may arrive faster but still need another check. Someone may have to find the missing information or ask the model to try again. Some errors may pass unnoticed. A cheaper model call does not remove any of that work.

That is what bothers me about the recent GPT releases. When I cannot see the reported improvements in my own use, I cannot assume that moving to the newer version is a good decision.

Onur Tırpan reports a similar problem in his tests

In his review of Opus 5.5 and GPT-6 Sol, Onur Tırpan describes tests drawn from his own business applications. The work involves more than producing a plausible answer. It includes finding and verifying relevant documents, extracting the necessary information, excluding irrelevant material, and preparing the results for the next step. These tasks depend on one another.

His reported findings echo my experience: some newer GPT models cost less and finish faster, but deliver less of the accuracy his application needs. These are not the same tests or model comparisons as the Terra report above. This is a practitioner’s account of his own applications; I have not independently examined the raw test data or all the settings used.

What caught my attention was his distinction between uncertainty that needs a human decision and an error that passes unnoticed. If a system recognizes an ambiguous case and asks a person, its checks may be working. If it omits information or invents something and proceeds as though the result were correct, the problem is different. An average score can obscure behaviors with very different consequences for a business.

He also reports that Opus 5.5 improved quality, cost, and speed together. His account does not support a claim that every cheaper model gets worse. What matters is how the replacement completes a particular task and which errors it makes.

Is lower cost coming at the expense of quality?

My experience with OpenAI’s recent releases sometimes leaves me with that impression. It feels as though some of the capabilities I need are being sacrificed as cost and speed improve.

But I cannot say what causes that impression. Seeing more errors during use does not reveal which technical decisions the provider made. Onur’s explanations about optimization and infrastructure are hypotheses too. They should not be presented as established causes of a decline in quality.

I do not know why I am encountering more errors with OpenAI’s latest models. But I do not need that technical explanation to choose a model. My criterion is how accurately it completes my work. If the previous model makes fewer errors on that work, continuing to use it makes more sense.

A new version has to earn its place in my work

In my article on why the same AI model produces different results across applications, I argued that context, tools, permissions, and workflow need to be considered alongside the model itself. An update changes one part of that system. We cannot assume that behavior which worked before will survive the change.

I also discussed how a cheaper answer can require more human review when choosing between models. This latest report brings me back to the same question: am I getting the same work for less money, or paying for the difference in correction effort?

There is no need to discard benchmark results. They need to be read with an understanding of the tasks and measures involved. But a claim of progress in an announcement cannot decide whether a business, or I, should change models.

A higher score than the previous model is interesting. If the replacement completes the same work with fewer errors, that gives me a reason to use it. With the last few GPT releases, I have not always seen those two improvements arrive together.