Which AI model for which task — 2026

One row per task

Each row of this chart is one of our leaderboards: text, coding, web development, agents, image and video. For every task you get three picks instead of a single winner, because the strongest model is rarely the one most people should pay for, and many readers want something they can download.

What the three columns mean

• Best — first place on that board. Ratings come from LMArena, where people compare two anonymous answers and vote.
• Best value — the cheapest model with a price among tiers S and A, measured by the blended price at its cheapest provider. If no S or A model has a price, the pick moves down to tier B.
• Open weights — the highest-rated model whose weights you can download and run on your own hardware.

Under each name you see the tier letter, the vendor and the price. Text boards are priced per 1M tokens of input and output; image models per picture and video per second. A click on a name opens the model in the side-by-side comparison, where you can put it next to the other two picks.

How to use it without being misled

First place is less decisive than it looks. Our tiers group models whose confidence intervals overlap, so the second and third model of a board are often statistically level with the leader and may be much cheaper. A rating also measures what arena voters preferred on their own prompts — mostly in English — not accuracy on your documents or your codebase. Treat the chart as a short list, then open the full board, read the votes and the interval, and try two or three candidates on a real task.

The open-weights pick is the starting point for running a model locally. Before you download it, check which open model fits your hardware: the memory it needs depends on its size and on the quantization you choose.

Yearly editions

Leaders change every few weeks, so the chart keeps one edition per year. Open a past year to see which models were new then and which ones have dropped out since.