Article

31 Oct 2026

The decision inside every AI request

Every AI request runs on a model and an effort level, and most teams never choose either. How to match both to the job without wasting money or quality.

Cartoon: a robot in a crane lowers a single sticky note to a bored worker while a tiny robot staggers under a tower of contracts

Pull up a chair beside almost anyone using AI at work and you'll find a small menu in the corner of the screen showing which model they're using and how hard it's set to think, still on whatever setting it had the day someone set up the account. Nobody chose it. A two-line email and a forty-page contract review both run through the same engine at the same effort, and when you point at the menu and ask about it, you tend to get a slightly guilty look, the one people usually reserve for terms and conditions they've never read. Underneath the look is a reasonable worry: that picking a smaller model or a lower setting means accepting worse work, and that nobody wants to be the person who economised on the job that went wrong.

So the default wins, every time, and the default is a decision somebody else made for you.

The choices hiding in a few ordinary jobs

When I tested Claude on a reference letter for a former employee, I ran it on Claude Sonnet at medium effort, a deliberately modest choice, because the letter was short, the context fitted comfortably into one reply, and the quality depended mostly on how well I answered the questions Claude asked me first.

For a much bigger job, I asked Claude to read a month of my email, cross-check it against our deals board, build an opportunity map in Miro and produce an infographic, all in one request. For that, I moved to Cowork and Claude's most capable model, Opus, because the job pulled in a large amount of context from four different systems and thoroughness mattered more than speed.

And when I asked Claude to turn a cartoon of me into a short animated video through Higgsfield, I named Kling 4.0 in the prompt. Some of the video models on that platform burn through credits far faster than others, and I had several videos to make that month.

None of those choices took more than a few seconds, and every one of them changed what the work cost and how good it turned out.

The setting beside the model menu

Next to the model menu in Claude is an effort setting, and plenty of regular users have never touched it. It controls how thoroughly Claude works through a task before answering, and at higher effort it thinks for longer, makes more tool calls and checks more of its own work. That extra care costs more of your usage allowance and takes longer, and the more capable models use your allowance faster too.

Anthropic's own guidance suggests trying lower effort on the more capable model before you drop to a smaller one to make your allowance last, because a frontier model at medium effort often outperforms an older one working at full stretch.

Too little effort, and Claude misses instructions or wraps up long jobs before they're finished. Too much, and you get longer answers that are no better, with the scope drifting past what you asked for (rather like asking a colleague for a quick view and receiving a fourteen-page memo with appendices). You've found the right level when Claude finishes the job in one or two turns. Anthropic suggests an easy way to find it: run a typical task at the default setting, then again one level lower, and compare what comes back. If you can't tell the difference, you've just found some free allowance.

The expensive mistake runs both ways

What I've found, watching teams settle into these tools, is that the costly mistakes come from both ends of the menu, and most businesses only ever guard against the obvious one. A premium model on maximum effort for a simple job wastes allowance and time, and sometimes produces worse work, buried in detail nobody needed. A budget model on a complex job looks cheaper on every run, right up until someone spends an hour fixing what it got wrong. Most of us have watched that second version play out, with the checking and correcting landing on whoever had the least time to spare that week. That person's hour costs more than any number of tokens.

On Higgsfield, eight seconds of Veo 3 video with audio costs around 58 credits, and its Fast version produces a similar clip for about 22. Across a dozen social clips a month, that gap comes to more than twice the monthly allowance on Higgsfield's entry plan. For a hero video on your homepage, the extra credits might be the best money you spend all quarter. For a throwaway internal clip, they're simply gone.

Count the cost of a finished piece of work, including the time someone spends checking and fixing it, and the cheapest run can easily turn into the most expensive outcome.

Defaults are policy

In most of the AI rollouts I've seen, nobody had a conversation about which model the team should use for what, even though that setting shapes quality and spend every day the team logs in. The default becomes the policy. Any leader who has had a Teams message on a Wednesday afternoon asking why Claude has suddenly stopped working, halfway through a client deliverable, has met that policy in person.

Write the team a one-page house rule. Everyday writing, summaries and quick questions run on the standard model at default or medium effort. Long, multi-step work that pulls information from several systems gets the most capable model. Anything that spends credits names its model explicitly and stays on approval until people know what a normal bill looks like. Put that page in your AI training, revisit it each quarter as the models change, and the team stops guessing. The most capable tools then stay available for the work that needs them, the analysis behind a board decision or the proposal that wins a contract, and nobody runs out of allowance by Thursday because they spent Monday asking a frontier model to fix their typos.

Somewhere in your business this week, a setting nobody chose will decide how good a piece of work is and what it costs. How many other decisions like that are running your AI programme right now without anyone's name against them?

Where to go next

If you would like help writing that house rule for your team, our AI Strategy Engagement is the place to start. For the other setting nobody chooses on purpose, read How much should your AI do without asking?.