You rarely need the biggest model.

“Should we just use the most powerful Artificial Intelligence (AI) model?” It sounds like the cautious choice. For most of the work an agency runs, it is the expensive one. A smaller model handles narrow, repeatable jobs at a sliver of the price, and the frontier earns its keep only where the task is genuinely hard.

What makes one model bigger than another?

These are all large language models (LLMs), and “size” points at two things at once: how many parameters a model carries (its raw capacity), and what it costs to run per word. A bigger model knows more and reasons further. It also charges more for every token² it reads and writes. The gap inside a single family is easy to see. Anthropic lists Claude Haiku, its smallest model, at one dollar per million input tokens, and Claude Opus, its largest, at five³. Same words, five times the price.

Have the smaller models caught up?

They have closed a surprising amount of ground. On the Massive Multitask Language Understanding benchmark (MMLU), a broad test of general knowledge, the smallest model to clear 60% in 2022 was PaLM, at 540 billion parameters. By 2024 Microsoft's Phi-3-mini hit the same mark with 3.8 billion, a 142-fold reduction⁴. The price of a fixed level of capability dropped just as sharply. Stanford's AI Index tracked the cost of GPT-3.5-level performance from twenty dollars per million tokens in late 2022 to seven cents by late 2024, more than a 280-fold fall in about eighteen months¹.

What's Real?

• •The cheap model wins the narrow, repeating work. Sort a support ticket, tag a photo, pull fields off a form, draft a first pass. NVIDIA's research group argues small models are sufficiently powerful, inherently more suitable, and necessarily more economical for many invocations⁵ in exactly this kind of pipeline.
• •The frontier still wins the hard part. Long chains of reasoning, open-ended judgment, the strange edge case a client will notice first. Paying for that capability where it changes the outcome is money well spent.
• •Mixing models is the normal design, not a hedge. The same NVIDIA paper calls heterogeneous systems that invoke several different models⁵ the natural choice. A Stanford method, FrugalGPT, matched the best single model with up to a 98% cost reduction⁶ by trying cheaper models first and escalating only when it had to.
• •You only learn “good enough” by testing. A benchmark headline is a starting point. Whether a small model clears the bar on your client's real work is an evals⁷ question, and it is answered on your own examples.

The “so what” — for anyone serving tech clients

• •Price the pipeline, not the demo. When you scope AI into a client's workflow, size each step on its own. The model that reads ten thousand product descriptions a day need not be the one that writes the launch strategy. At volume, that difference is most of the bill.
• •Ask your vendors which model runs where. “We use AI” is not an answer. A partner who can say the small model handles the routine calls and the frontier takes the judgment calls is a partner who has thought about your cost.
• •Don't overpay for reasoning either. Putting a heavy reasoning model⁸ on a simple lookup is the same mistake wearing a different hat. Match the tool to the task, every time.