AI has created a new optimization problem

One of the most interesting patterns we are seeing at Pensero has nothing to do with AI adoption itself but with how people choose which models to use.

Across the engineering organizations we benchmark, and even within our own team, there seems to be a natural tendency to default to the newest and most expensive model available.

If someone has access to Claude Fable, GPT-5, Gemini, or another frontier model, they ted to use it for almost everything and not just the complicated tasks.

What is becoming the online joke: let’s use Fable to center a div.

The right tool depends on the job

You do not need a Lamborghini to do your grocery shopping.

• Could you drive one to the supermarket? Of course.

• Will you arrive? Absolutely.

• Does buying milk require a supercar? Not really.

So the same applies to AI.

There are tasks that genuinely benefit from the strongest reasoning models. Designing a complex system architecture, debugging a difficult production issue, reviewing a critical technical design, or exploring several possible implementation strategies are all good examples.

There are many more that do not.

Writing documentation, summarizing meeting notes, cleaning up code, generating unit tests, formatting data, answering straightforward questions, or making small code changes can often be completed with significantly fewer tokens and at a fraction of the cost.

Using the most capable model for every prompt is not a sign of optimization. On the contrary, it is often a sign that no optimization is happening at all.

AI spend is becoming an engineering problem

This matters because AI is no longer an experiment.

Engineering organizations are now spending real money on inference. Companies are rolling out Cursor, GitHub Copilot, Claude Code, Gemini, ChatGPT, and a growing number of AI agents across hundreds or even thousands of engineers.

The conversation is gradually moving away from, "Should we adopt AI?". The new question is, "How do we generate the highest return from the AI we are already paying for?"

That is a much more interesting problem.

The best teams think economically

Looking across thousands of engineers, one pattern becomes increasingly clear: The organizations generating the most value from AI are not necessarily the ones consuming the most tokens or deploying the most powerful models. They are the ones that have developed better decision-making habits around AI.

They understand that premium reasoning should be reserved for the problems where it genuinely changes the outcome, while simpler tasks should be handled by models that deliver the same result at a lower cost.

Those small decisions compound over time. Saving a few cents on a single prompt is irrelevant, but making millions of those decisions across hundreds or thousands of engineers eventually becomes a meaningful competitive advantage.

Engineering has always been about trade-offs

Good engineers rarely optimize for maximum compute, maximum infrastructure, or maximum complexity. They optimize for the outcome and so AI should be treated the same way.

As AI becomes another layer of engineering infrastructure, the teams that create the most value will not be those consuming the most tokens or defaulting to the most powerful model every time. They will be the ones that understand when additional intelligence changes the outcome and when it simply increases the invoice.

That discipline is becoming a competitive advantage. As AI usage scales across organizations, every unnecessary prompt to a premium model becomes another avoidable cost. Every task matched to the right model improves the economics of the entire engineering organization.

Just as you do not need a Lamborghini to buy groceries, you do not need your most expensive AI model to center a div. The goal is not to use more intelligence. The goal is to use exactly as much intelligence as the problem requires.