The whole point of OpenAI's Responses API is to help them hide reasoning traces (www.seangoedecke.com)

OpenAI’s new Responses API, released six months ago to replace the simpler /chat/completions endpoint, is designed to be stateful, allowing the backend to manage conversation history instead of requiring clients to handle it. While OpenAI highlights performance and cost benefits, as well as added agentic capabilities unlocked by this API, the deeper motivation is to keep their models’ reasoning traces—internal chains of thought—hidden from users. Unlike competitors like Claude or Qwen, which expose these reasoning processes in their API outputs, OpenAI’s latest models (e.g., GPT-5-Thinking) do not reveal such chains, likely to protect proprietary methods or sensitive information. This secrecy creates a challenge: without access to these internal reasoning steps, developers using the stateless /chat/completions API cannot preserve or leverage the chain of thought, resulting in less capable third-party applications. The Responses API circumvents this by maintaining the reasoning traces internally on OpenAI’s servers, integrating them seamlessly into conversations but never exposing them to clients. This design choice lets developers harness the full power of OpenAI’s reasoning models, but it comes with trade-offs in transparency and user control. The Responses API’s positioning as a more flexible, cost-effective alternative obscures the core issue—it exists primarily to work around OpenAI’s decision to conceal their models’ internal reasoning. This contrasts with competitors like Anthropic, which continue to support stateless APIs that openly share chain-of-thought details. For the AI/ML community, this raises important questions about transparency, user agency, and the trade-offs between model performance and openness in AI deployment.