Anthropic blocked Claude Code subscription usage for OpenClaw, forcing a search for alternative LLMs. After finding GPT-5.4 too lazy and unreliable for agentic tasks, the author tested Chinese open-source models and settled on Kimi-K2.5 via OpenRouter. Setup involves fetching an OpenRouter API key, removing all Anthropic environment variable references to avoid OAuth issues, and applying standard OpenClaw optimization tips (task-specific skills, permissions, cron jobs for learning). Kimi-K2.5 ranks just below Claude Opus 4.6 in performance but costs roughly 1/10th the price. Key downsides include slower response times on simple queries due to excessive thinking tokens, and GDPR non-compliance when using the hosted API — though self-hosting is possible since the model is open source.
Table of contents
Why use OpenClaw with open-source modelsHow to use Kimi-K2.5 in OpenClawKimi-K2.5 PerformanceTechniques to optimize OpenClawDownsides of Kimi-K2.5ConclusionSort: