Why I don't chase every new AI model
Writing a better email doesn't require a new generation of artificial intelligence. My take on the model madness and when an upgrade actually helps.
By the time you've remembered a model's name, it's already old. New version, new benchmarks, fresh excitement in your feed. Somewhere, someone is explaining why the model released three weeks ago is now completely useless, obviously.
Sorry, but have we lost the plot?
I develop applications with AI myself, and I find the technical progress fascinating. Still, the constant upgrade circus is getting on my nerves. Especially when every new release is treated as something everybody suddenly needs.
A shopping list doesn't need a flagship
Making an email friendlier. Summarising a long text. Sorting out an idea. Explaining an Excel formula. These are the kinds of tasks many people use AI for every day.
My opinion is straightforward: By the time we reached the level of GPT-5.6, we already had more than enough capability for a large share of these tasks. If I had to set a practical upper limit for ordinary everyday use, that would be it. Often, I'd choose something below it.
The next generation doesn't need to understand my shopping list on a deeper level.
Here, GPT-5.6 means the capability level of the model family, rather than one particular configuration everyone should use. At its July 2026 launch, OpenAI already distinguished between Sol, Terra and Luna: a flagship, a balanced option for everyday work and a particularly cost-efficient option. There are good reasons to choose a smaller option even within one generation.
That's my assessment of typical everyday tasks, not a claim that every problem is solved or every error has disappeared. A wrong answer is still wrong, however confidently the model delivers it.
Three weeks later, everything is new again
The pace is real enough. Google released Gemini 3.7 Flash on August 13, 2026, just three weeks after Gemini 3.6 Flash according to its announcement.
For developers, that can be worth investigating. Better results, fewer failed attempts or lower costs can meaningfully improve an application. I intend to keep following these releases.
What bothers me is the automatic response: It's new, so switch immediately. A release is an option. Whether I need it depends on the work I'm doing.
If my existing setup works reliably, I don't need to replace it out of nervousness. A version number isn't there to keep me busy.
Progress means more than a bigger number
A newer model doesn't necessarily have to be larger or more expensive. Progress can mean completing the same task faster, more cheaply or with fewer corrections. That does interest me.
For someone editing a few texts a day, a slightly higher benchmark score isn't much of a reason to switch. Will I notice the difference? Will it save me time? Will I have less to fix?
Anthropic describes model selection as a balance between capabilities, speed and cost, and recommends testing with your own tasks. Rather sensible, really. Almost suspiciously sensible for an industry that seems to announce the next big thing every week.
Where I do want more capable models
Scientific questions, complex software engineering and analysing large collections of conflicting documents are a different matter. Extra capability can make the difference between a useful result and an elegantly worded dead end.
A major refactoring across many files is a different task from writing a birthday message. A difficult research problem deserves more effort than figuring out what to cook with three carrots and a courgette.
As a developer, I'm curious about more capable models. I want to see which previously difficult tasks they make possible. That doesn't oblige me to use them for every little job.
And the task matters more than the label “ordinary consumer”. Someone who rarely uses AI can still have a demanding question. By all means, bring in the stronger model then.
My criteria are fairly boring
To judge an upgrade, I'd give the existing model and the new one the same concrete tasks. With the same information and clear criteria:
- Is the result correct and complete?
- How much do I have to fix myself?
- How long does it take to reach a useful result?
- What does the completed task cost, including retries?
If the new model is noticeably better, there's a reason to switch. If it only looks better in my feed, I can save myself the trouble.
An upcoming retirement or the end of support is another valid reason, of course. The point is to make a deliberate decision and plan ahead.
A little less AI madness, please
I'd like more attention on what we can usefully do with the models we already have. Better applications. Understandable interfaces. Good information to work from. Results we can check.
For many everyday tasks, the models are already capable enough. We still need to build something useful around them.
I'll stay curious. But I'm done chasing every version number. My emails can be friendly enough without the latest breakthrough.