The most expensive model for everything is the most expensive mistake.

When a company introduces AI, it naturally reaches for the most powerful model. It's the best in the tests, so why not. But the most powerful model sorting emails is like a law firm partner sorting the mail. They can do it, but they're slow, expensive, and their time is missing elsewhere.

Not three technologies, but three types of workers

All the models we talk about here are language models. They differ in size, and therefore in what they can do and what they cost. They fall into three groups:

• Small models (technically SLM, small language model). Specialists in one clear task. Fast, cheap, and they can even run on your own server.
• Large models (LLM, large language model). Generalists with a broad view. They handle conversation, subtle differences in meaning and situations they haven't seen before.
• Frontier models. The most capable today, for example Claude Opus, GPT or Gemini Pro. They reason best across long chains of steps and use tools on their own.

Which model stands where: a comparison as of October 2026

Leaderboards change every month, so we'll update this overview regularly. The chart shows what a model can do and what it costs. Each curve is one model at different levels of reasoning: the longer it thinks, the more capable it is, but also the more expensive. Notice that some cheaper models get very close to the most expensive ones.

A typical task for a small model is sorting. Hundreds of documents arrive at a company every day, and each has to be classified and sent to the right person. It's pattern recognition, not deep thinking. A well-⁠tuned small model often handles it as well as a large one, only faster and cheaper.

The second advantage is data sensitivity. A small model can run on your own server. Documents never leave the company, and the question of where patient or client data goes disappears.

The difference in price and speed is large. According to Artificial Analysis, Claude Haiku 5.5 costs about $0.41 per test task, while Claude Opus 5.5 costs over $10, roughly 25 times more. Haiku also responds more than twice as fast, 241 versus 98 tokens per second. With thousands of emails a month, that's a difference you'll feel.

If data must not leave the company, you can use open-⁠weight models that run on your own server, for example the DeepSeek or Xiaomi MiMo families. But they require your own hardware and someone to look after them.

A different situation is a customer complaining about an invoice. The problem relates to a service change two months ago and three older complaints. This requires combining data from several places and understanding how it connects. And every customer describes the same problem in different words.

That's what a large model is for. It has seen far more different situations in training, so it can handle a combination it hasn't met before.

The gap between a large and a frontier model is often smaller than you'd expect. GPT-6.1 Sol scores 55 on the index at $1.27 per task. The frontier GPT-6 Astra scores 57, but for $6.24. Two extra points cost almost five times as much. For customer replies or summaries, you usually won't notice the difference. On the invoice, you will.

The most powerful models make sense where a task has no clear procedure. You need to find the cause, plan a fix, carry it out across several systems and keep checking that it works. A typical example is a system outage at two in the morning that would otherwise wake up a technician.

But reality is more sober: most companies today don't let such an agent work entirely on its own. It works as an assistant with rules, and a person has the final say. Still, the capabilities it needs belong to the most powerful model.

Even with a frontier model, the setting matters. At a medium reasoning level, Claude Opus 5.5 scores 48 points on the Artificial Analysis index for $2.45 per task. At the maximum level it scores 58, but for $10.04. The same model can be a reasonably priced assistant or an expensive expert. It depends on how you set it up for the task.

Credits without a cap are a symptom, not the cause

Companies often ask us why their AI bill grows faster than expected. The cause is usually the same: one expensive model does everything, from sorting to complex analysis. The second cause is that nobody set limits.

That's why we make two decisions in every solution. Which model for which task. And what monthly usage cap, so the invoice doesn't surprise you. We bill model credits by actual usage, with a set limit.

Choose the model by the task, not by the leaderboard

Model leaderboards change every month. Your tasks change much more slowly. If you build the solution so the model can be swapped, you don't depend on one vendor or its price list.

When we start with a company, we first write down the tasks AI should do. Only then do we assign a model to each. The result is usually a combination of two or three models, not one for everything.

Jaro Rais

Founder of itashu. Builds agents and connects AI to company systems.