mg. Mauricio
Giacometti Let’s talk
AN EXPERIMENT / LANGUAGE MODELSARTIFICIAL ANALYSIS DATA · 2026-10-19

1 · What it is for

Which model should
my agent use?

If you are building an agent, you have to choose the language model that runs it. There are dozens, and they all claim to be good. This page helps you decide in two steps: first, understand how the market lines up; then, if you want, find the model that fits your task and your budget.

There is no “best model”

The one that wins at writing code is not necessarily the one that wins at reading documents. It depends on the task and on how much you can pay.

Capability is expensive

Between the most basic tier and the most capable one, reasoning capability grows 1.9 times and the input price grows 10 times. Doubling capability costs five times more.

“Fast” means three things

How long it takes to start, how many tokens it writes per second, and how long it takes to finish a task. A model can write fast and still take minutes to begin.

Basis:

See the map ↓

2 · How the market lines up

The map of the 24 models

Each point is a model. What matters is where it sits:

  • To the right (PC1): more capable. One axis summarizes four quality tests: reasoning, coding, agentic work and documents.
  • Up (PC2): writes more tokens per second once it has started. It is not response time: a model near the top can take minutes to begin.
  • Color and shape: the tier. Size: how much it costs. Bigger means more expensive (estimated weighted cost per million tokens).

Hover over a point to see its data; click it to see its profile.

PC1 and PC2 are principal components, not variables: PC1 explains 70.1% of the variance and PC2 13.6%. Latency was left out of the model. Source: artificialanalysis.ai/models, cutoff 2026-10-19.

Four tiers of the same ladder

We group the models into four tiers so we can talk about them. They are not four kinds of model: they are four cuts of a continuous ladder where capability rises, speed falls and price follows. The split is not validated: nobody tested whether it holds when the data changes.

Why “fast” is misleading: the three speeds by tier
Averages by tier. Source: artificialanalysis.ai/models, cutoff 2026-10-19.
MeasureC0C1C2C3
Output throughput (t/s)508.6122.751.2110.7
Latency to first token (s)1.441.2114.0304.1
Time per task (s)805981,106703

The frontier tier writes fast but starts late, because it reasons before emitting. The heavy-lifting tier starts earlier but writes at half the throughput, and ends up the slowest end to end.

3 · Pick your model

Now: the best one for your case

If you have seen how the market lines up, here you can go into detail: choose the task, set your limits on price, context and waiting time, and see which models stay on top and which give more for less.

Open the explorer: filters, ranking, cost and full table
  • K2 Horizon (†) has an imputed price and is not sold commercially. The source publishes 0.00; it was assigned the dataset minimum. If it tops a cost ranking, it is not an option you can buy.
  • Tasks are not mixed. Each one is a different test; no average across them is computed.
  • “No data” is not a low score. It means the test was not run for that model.

Ranking

Model profile

Cost versus capability

X axis: the score on the selected task. Y axis: cost, on a log scale. The line joins the Pareto frontier: the models for which no other model is both better and cheaper.

All metrics

Click a header to sort. Missing values are shown as “no data” and always go last, in either sort direction.

What this page cannot tell you

Data: artificialanalysis.ai/models, cutoff 2026-10-19. Analysis and tool: Mauricio Giacometti.

Code and data: github.com/mgiacomettiferrer/que-modelo-elijo