15Free and cheap

Good news: free works, and these aren't toys anymore. In summer 2026 Kimi K3 launched straight into the top three models in the world on the Artificial Analysis ranking, and DeepSeek and Qwen are breathing down the leaders' necks. Less good news: for long agent work the paid Claude and Codex are still more reliable, and free tiers squeeze you with limits.

That doesn't mean "don't use them". It means use them knowing what you're using.

Qwen - from Alibaba. Does a bit of everything, from code to images; probably the most all-round of the free ones.

Kimi - from Moonshot, that same K3. Strong on code and long analysis.

DeepSeek - strong on code, and through the API the prices look like a typo next to everyone else's.

Z.ai - the GLM models. The chat is free, and for code there's a GLM Coding Plan subscription from eighteen dollars that plugs straight into Claude Code or OpenCode.

MiMo - a fresh model from Xiaomi, free to try right on the site.

Now the downsides. Limits are tighter than on paid plans. On long, complex tasks they lose the thread more often. Some are weaker in languages other than English and Chinese. And you can pair them with agents like Claude Code, but expect to fiddle with the setup.

How I'd use them: learn, experiment, do small things. And when you hit the ceiling - that's the moment paying starts to make sense.