The hubris is strong in this one. Straight from an agent’s thinking block while wiring up an LLM call: “There’s no existing Anthropic key block, only an openai-api-key one. I’m weighing options … I’m deciding to go with Anthropic and Claude since model quality matters for this judgement task.”

Note that the existing OpenAI key would have worked immediately. Consider this your daily reminder not to trust your model blindly, especially when it is reasoning about itself.