OpenAI recently released GPT-5.6, claiming it delivers higher performance at a lower cost than the previous generation. API prices have dropped again, and the amount of work that fits within the same budget has grown. Faced with this news, many practitioners' minds jump to the same question: is it time to switch subscription plans?
The question itself is natural. But there's something worth checking before you answer it.
Once the Tools Level Off
In the early 1990s, once spreadsheet software had spread through most offices, simply knowing how to use one stopped being a differentiator before long. What created a gap in practice after that wasn't whether you'd adopted the software. It was judgment — what problem you solved with the same tool, how you solved it, and where you started looking when a number came out wrong.
AI model prices have fallen fast over the past two years. Accessing a high-performance model in 2023 cost tens of times more than a comparably capable model does today. This GPT-5.6 announcement is simply the latest chapter in that trend. It's not just OpenAI, either — Google, Anthropic, Meta, and the other major providers are all racing in the same direction. If this pace holds, the window in which "which model you use" functions as a real barrier to entry keeps shrinking.
There's a common expectation that as AI gets cheaper, gaps between people narrow. That's not wrong, in the sense that more people gain access to better tools. But for that to actually play out, one condition has to hold: that the difference in ability among the people using the tool isn't very large to begin with. Whether that assumption actually holds is a separate question.
What Splits Apart on the Same Model
Among the variables that determine the quality of AI output, model choice matters less than you'd think. Performance differences between models certainly exist — in controlled tests, newer models consistently beat older ones. The problem is that a working environment is not a controlled test.
Feed the same model the same prompt, and one person will use the result as-is, another will cross-check the key figures against an independent source, and a third will catch the premise missing from the logic. The gap between the final outputs these three paths produce is bigger than the gap between model versions.
The stronger the model gets, the sharper this divide becomes. The more polished the output looks, the stronger the pull to accept it at face value. Grammar errors vanish, sentences read more smoothly, and the logic flows more naturally — which means errors now arrive wrapped in more persuasive language.
In 2023, a lawyer in a New York court filed a legal brief drafted with ChatGPT that cited six court cases that didn't actually exist. The filing went in as the AI had produced it, and the fabrication only surfaced when the court tried to verify the citations itself. This isn't a problem confined to law. The more fluent AI output becomes, the more often it slides through without a second look — and in that environment, the cost paid by someone who hasn't built the habit of verification only rises as the models get better.
The Capability That Builds Outside the Tool
Knowing what to hand off to AI and what to hold onto yourself isn't a skill that develops automatically just because you spend more hours using AI. When you genuinely know a field, you can immediately sense whether a conclusion falls within the range of common sense. What a given customer responds to, which line of reasoning actually holds up in practice — none of that lives inside the text AI processes. It's built from real experience.
The range of what AI can handle keeps expanding. In drafting, summarizing, classifying, and writing code, the speed gap between AI and people has already widened considerably. Once you accept that, the question changes. It stops being "which model is better" and becomes "what am I adding to what the AI produces."
People who can answer that second question benefit as AI gets cheaper — the same judgment now lets them handle more work at a lower cost. For everyone else, a stronger model doesn't actually raise the quality of the output. It just produces roughly the same result faster, and more convincingly.
If a large share of what comes out of the model you're using today goes out the door unedited, upgrading to a better model — until that gap closes — only deepens your dependence on it. Before you check your AI subscription bill, check how closely you're actually reading what your current model produces.
