PlainVerdict
Menu

Models, apps and harnesses: what are you actually using?

The AI model, the app you chat in and the tool wrapped around it are different things. What open weights means, and what our data can and cannot say about free access.

This guide explains general ideas and uses facts from our datasets where it can. It contains no prices, because we have not verified any.

The basics, and what our data shows

Four things people mix up

  • The model is the trained AI itself, such as the names in our table.
  • The app or website is where you chat with it. One app may offer several models, and the same model can feel different in different apps because the app adds its own instructions and settings.
  • The API is how developers connect a model to their own products. It is billed by use. See how AI pricing works.
  • A harness is the tool wrapped around a model to do a job, such as a coding assistant that reads your files and runs commands. The harness changes the result, not only the model.

What our data shows about this

  • The FrontierCode coding results we publish were measured with these tools: chisel, claude-code, codex, devin-cli, grok-build and mini-swe-agent. Because the tools differ, we show those scores but never rate them.
  • The models were tested at these settings: 0110, 0309, 20251101, 26-02-10, High, High, 20260813, Max, Medium, NVFP4, Reasoning, 0309, Thinking, high, low, max, medium, none, xHigh and xhigh. A higher setting generally means more thinking before answering, and the sources do not always test the same setting for the same model.
  • 20 of the 77 listed models carry an open licence in Arena's data or an open-weights label from Epoch AI, from these makers: Alibaba, DeepSeek, Google DeepMind, Moonshot, NVIDIA, Thinking Machines, Xiaomi and Z.ai (Zhipu AI).
  • DeepSeek's listed models and their licences: DeepSeek-V4-Pro (MIT); DeepSeek-V4-Pro (High, 20260813) (MIT); DeepSeek-V4-Pro Preview (MIT); DeepSeek-V4.1-Flash-20260910 (MIT).

What “open weights” means

The model's files are published under a licence that lets other people run it. That means it can be offered by more than one company, and some people run it themselves. It does not tell you what any particular app charges.

What this means for you

When people say a model is “pretty much free”, they are usually talking about one app's free plan, or about an open model that someone hosts cheaply. Those are different claims, and our data covers neither: it shows the licence, not the price or the free allowance of any app. We have not verified any free tier, so this site does not make that claim for DeepSeek or anyone else.

The practical question is: which app or tool will I actually use, which model does it run, and what are the limits and costs of that plan? Our ratings tell you how the model did on tests. They do not tell you what you get in a given app.

What to do next

Pick your job in Find my AI, then check the maker's own page or the app you plan to use for which model it runs, what the free plan includes, and what a paid plan adds. You can filter to open-weight models to see the ones that can be offered by more than one provider.