A YouTube video about Claude Code stuck with me. The creator said they set the model to Anthropic’s Opus and never changed it. Opus for everything: simple fixes, big refactors, everything in between.
That made me wince. It is the software equivalent of bringing a 1000cc sportbike to a go-kart track.
The Wrong Track for the Wrong Bike
I have spent much time at track days, mostly on smaller bikes. The best show is a rider on a liter bike opening the throttle on the straight and blasting past a 300cc bike. For a few seconds, power looks like the winner. Then the corner arrives. The big bike has to slow down, and the rider on the 300cc bike slips past on the outside, already set up for the next turn.
Experienced riders on 300cc bikes often run faster, more consistent laps than inexperienced riders on liter bikes. They are more efficient on a machine with much less power.
Skill Beats Displacement
A liter bike does not make you faster if you do not know when to brake, how to trail-brake into a corner, or when to pick up the throttle. The 300cc bike lacks top-end speed, but it rewards good technique. The machine does not close the gap. The rider does.
I have watched coaches ride around faster bikes with their visor up, gesturing with one hand, while the person on the bigger bike is holding on for dear life. The smaller bike costs less to buy, less to fuel, and less to maintain. It also asks less of the rider.
The Same Pattern in AI Models
People who default to Opus for every task are doing the same thing. Anthropic’s Opus is a powerful model, and some problems need that power. Using it for every small job is like opening the throttle on a straight and then not knowing how to handle the corner. You can move fast. You can also crash fast.
A cheaper model, or even a free one like Cognition’s SWE-1.6, can be the right tool when the task is small and the operator knows what they are doing. The savings are not just the per-token price. A smaller model stays focused, complicates things less, and is easier to keep on track. The total cost includes the time it takes to clean up an overly elaborate response.
Cost Is More Than the Price Tag
When I think about the liter bike, I think about tires. A 1000cc bike eats expensive tires faster than a 300cc bike because of the power it puts down. The same thing happens with models. Running the most expensive option for every prompt burns tokens like a liter bike burns rubber. The bill is one problem. The wasted motion is another.
Match the model to the job. I get more done with good prompts on a cheap model than with mediocre prompts on a premium one.
Picking the Right Machine
I am not against powerful models. When I need deep reasoning or a long context, I reach for Opus or something like it. I choose deliberately, the way I choose a bike for a specific track. A tight, technical course favors a light, agile machine. A wide, fast track might call for more displacement.
The discipline is the same: know the terrain, know your own skill, and pick the tool that lets you be consistent.
What I am learning is that the model is only one part of the system. The operator matters at least as much. A smaller model in practiced hands can outrun a flagship model in uncertain ones.
Leave a Reply