Anthropic's S-1 says open-weight models could win some workloads and pressure its prices. That doesn't mean Claude is losing. It does mean your agent workflow sits on commercial decisions you don't control, so a backup model is worth having.

The catch is that "backup" is usually just an endpoint. DeepSeek and Kimi both speak the Anthropic format, so pointing Claude Code at them takes minutes. Whether the agent still works is a different question. In one reported case, a small local Qwen model behind Ollama returned tool calls as plain text JSON instead of tool_use blocks, four times running. Valid responses, zero finished tasks.

Three checks before you trust a fallback:

  1. The protocol carries tool calls, tool results and streaming intact.
  2. Your real harness still enforces hooks and permissions on a throwaway repo. The harness does that, not the model.
  3. About five real tasks, several runs each, with the patches actually validated.

Then compare cost per accepted task, including failed attempts and review time, instead of price per token.

Full breakdown, including the pricing math and the routing traps: https://levelup.gitconnected.com/anthropics-s-1-names-its-real-rival-cheap-open-weight-models-05d11f23d229?sk=d17020566f668a2d02482e5f20461815