← Todos los artículos

How to choose an AI model in 2026: cost per finished task, not per token

Choose an AI model by its cost per correct result on your own tasks, not by its price per token or its place on a leaderboard. Start with a capable model, build a test set that the task must pass, then lower the cost one step at a time, first by reducing the model's reasoning effort and only then by moving to a cheaper tier, and repeat the check whenever a provider releases a new model. In my own study of 736 agent sessions, the jump from a small local model to a cheap hosted one mattered more than anything else. Above that point, a flagship costing several times more scored no better.

What should decide the choice of an AI model?

Six things should decide the choice of an AI model, in roughly this order:

  1. The task: classification and extraction, long agentic work, hard reasoning, code or writing each favour different models. How these models work, and what none of them can do, is explained in what a large language model is.
  2. The quality bar: the score on your own test set that the task must reach to be useful.
  3. Cost per finished task: tokens times price, plus retries and the cost of failures, measured on real cases, never estimated from the price list alone. The parts of that bill are explained in what an AI agent costs to run.
  4. Latency: whether a person is waiting for the answer.
  5. Data rules: where the data may be processed and kept, which can rule out a provider before any test, as covered in GDPR and AI.
  6. Language: for Greek work, cost and quality must be measured on Greek text, because both differ from English.

How much better is an expensive AI model?

An expensive AI model is much better than a small local one and often no better than a mid-priced hosted one. In How much of an agent is the model, I ran six agent harnesses on a ladder of models, with every score expressed as a share of the room above an agent that does nothing:

  • a small local 8-billion-parameter model scored 0.142 and completed 17% of coding tasks;
  • a cheap hosted model scored 0.719;
  • a flagship model scored 0.788, at several times the cost;
  • a newer model at maximum reasoning scored 0.767 while running faster and cheaper than the flagship, and completed 94% of coding tasks.

The ladder is steep at the bottom and almost flat at the top. For most business tasks the decisive step is leaving the small local model, not reaching the most expensive one.

When is a small or local model enough?

A small or local model is enough for narrow, high-volume tasks with short, checkable outputs: classifying messages, extracting fields, routing requests, flagging duplicates. It is not enough for long agentic work, hard reasoning or writing code at scale, where the pass rate sets the bill. Small models can also be trained for one narrow job. The 5-million-parameter intent classifier I built runs in a browser tab, which is the right shape for a single, well-defined decision and the wrong shape for anything open-ended. How to design a narrow decision so a small or specialised model can take it is in Jev and structured outputs.

Should you use an open-weight model or an API model?

Use an open-weight model when you have a reason the API cannot satisfy: data that may not leave your infrastructure, a steady high volume where a reserved machine beats per-token pricing, or a latency you must control. Open-weight models from several labs are competitive in 2026 and many carry permissive licences, though some, such as Llama's community licence, are not open source in the strict sense (Open Source Initiative), so read the licence before shipping. Do not self-host to save money on bursty or low-volume traffic, because an idle GPU costs the same as a busy one, and the API provider absorbs that idle time for you.

Should you use several AI models at once?

Use several AI models at once only when each has a fixed job, such as a small model classifying and a large model writing. Dynamic routing, where software guesses per request whether a cheap model will do, looks attractive and disappoints in practice: research routers recover only part of the gap to a perfect choice, and every switch between models forfeits the prompt cache, because caches belong to one model. Before building a router, measure the same model at a lower reasoning effort. It keeps one cache, one prompt and one set of behaviours, and providers themselves recommend pricing it first (Anthropic, choosing a model).

How do you avoid being locked into one AI provider?

You avoid lock-in by keeping the provider a configuration value rather than a design assumption, and accepting that switching still costs work. Keep the model name in configuration, keep the conversation state in your own system, define tools in plain JSON Schema, and translate each provider's usage figures and stop reasons into your own fields. Most importantly, make the test set run against any provider. Prompts do not port cleanly: both major providers say a prompt tuned for the other underperforms, so a switch means re-running the tests and re-tuning, not changing one address.

What is a practical procedure for choosing a model?

A practical procedure for choosing a model takes five steps:

  1. Build a test set of 30 to 100 real cases for the task, with a grader, as described in LLM evals.
  2. Run a capable model at a normal setting and confirm the task passes at all.
  3. Lower the reasoning effort, then try a cheaper model, one change at a time, recording quality and cost per finished task.
  4. Remove any option that fails the data rules or the latency need.
  5. Re-run the set whenever a provider releases a model or changes prices, because the best choice moves every few months.

This is how I choose the model for the AI agents for business I build.

Frequently asked questions

Which is the best AI model in 2026?

There is no single best model. The best model for a task is the one that passes your own test set at the lowest cost per finished task, and that answer changes with each release.

Is ChatGPT or Claude better for business?

Both families are strong, and the answer depends on the task, the data rules and your test results. I use models from OpenAI and Anthropic in production and pick per task after measuring, not by brand.

Does the choice of model matter for Greek?

Yes. Models differ in how naturally they write Greek and in how many tokens Greek text costs them, as measured in Greek in AI models. Test on your own Greek material before choosing.

If you are choosing a model for a real workload, send me the task and a sample of the inputs. A short measured comparison beats a month of reading leaderboards.