They should be concerned with open weight models that don’t run on consumer hardware. The larger models from Qwen (Qwen Max) and ZLM (GLM and GLM air) perform not too far from Claude Sonnet 4 and GPT-5. ZLM offers a $3 plan that is decently generous. I can pretty much replace it over Sonnet 4 in Claude Code (I swear, Anthropic has been nerfing Sonnet 4 for people on the Pro plan).

You can run Qwen3-coder for free upto 1000 requests a day. Admittedly not state of the art but works as good of 5o-mini