Stop Paying for the Big Model

You are using the most expensive model for everything. Every draft, every summary, every quick question goes through the same heavyweight that costs 30 times more than its smaller sibling. And most of the time, the answer is identical.

This is not about quality. It is about knowing which tasks need the big model and which ones do not. The difference between getting this right and getting it wrong is the difference between a $3 monthly bill and a $90 one.

Four rules to stop overpaying without losing quality.

1. Start with the smallest model that could possibly work

I need to [describe the task]. Give me the result using the simplest reasoning possible. If the task requires deep analysis, say so instead of guessing.

Most people start at the top and never come down. The right approach is the opposite: start with nano or mini. If the output is good enough, you are done. If it is not, you know exactly what the cheaper model missed, and you move up one tier with that information.

GPT-5.6 nano costs $0.15 per million tokens. GPT-5.6 standard costs $5. Claude Haiku costs a fraction of Opus. The gap is not small — it is two orders of magnitude. And for classification, extraction, summarisation, and formatting, the small model gets it right more often than people expect.

2. Use the big model only for judgement calls

Review this [draft / analysis / decision]. What is wrong with it? What is missing? What would change your recommendation?

The big model earns its cost when it has to weigh tradeoffs, catch subtle errors, or make a call that requires understanding context across a long document. That is judgement, not computation.

A useful pattern: let the small model generate, then let the big model review. The generation is cheap. The review catches what matters. You pay the premium price only for the premium skill.

3. Batch the cheap work, isolate the expensive work

Here are 20 customer support emails. For each one, classify the intent, extract the product name, and flag urgency (high / medium / low). Return a JSON array.

If you are running the same operation on 50 inputs, a single batch call to the small model costs almost nothing. Running them one by one through the big model costs 50 times more and takes longer.

The rule: if the task has a clear right answer and the input is structured, batch it on the small model. If the task requires reasoning about ambiguity, send it to the big model — but only that task, not the whole pipeline.

4. Measure before you assume

Run this same prompt on [model A] and [model B]. Compare the outputs side by side. Where do they differ? Which differences matter for my use case?

The most expensive mistake is assuming the big model is better without checking. Run the same prompt on both. Read the outputs. Half the time, they are functionally identical. The other half, you learn exactly where the big model adds value — and you can route only those cases to it.

This is not a one-time exercise. Models update. Prices change. What needed the big model six months ago might not need it today.

The real cost is not the model — it is the default

The problem is never that someone chose the big model for a hard task. The problem is that someone set the big model as the default and never looked at the bill.

Pick the smallest model that works. Move up only when you see it fail. Measure the difference. The money you save is not a rounding error — it is the difference between a side project you can afford and one you cannot.

Want the calm version of AI news like this, once a week? Subscribe to the Sharp AI Hub newsletter →