Moving to the Responses API: A Migration Checklist

Look forGPT-6 Astra and GPT-6.1 Sol support Chat Completions, and they call tools only through Responses. GPT-6 Sol and GPT-6 Luna support function calling in Chat Completions only with reasoning effort set to none.

Look forReplace messages with input and move response_format to text.format. With reasoning effort above none, remove temperature, top_p, and logprobs. On GPT-6, replace prompt_cache_retention with prompt_cache_options.ttl.

Look forOpenAI stores responses by default, so set store to false if you keep history yourself. Assistants moved to the Responses and Conversations APIs. Conversations keep state until you delete it, and Zero Data Retention does not cover them.

Look forUse Responses when your application controls the loop. The Agents SDK adds handoffs and tracing in your runtime. The beta Agents API runs Codex's agent loop as a service OpenAI manages. Since October 9, 2026, its hosted sandboxes bill compute only during active turns. They can be pre-warmed and support computer use in an OpenAI-hosted browser. One builder moved every project to the SDK, while another moved a client to a different framework.

Look forOpenAI reports that Responses makes 40% to 80% better use of the cache than Chat Completions in its internal tests. Confirm it on your own traffic, and keep tool lists and tool_choice stable so the prefix does not change.

Look forMigration work sized from your existing state and retrieval code. A retrieval-augmented generation (RAG) chat team serving tens of thousands of users moved from Assistants to Responses in about two hours. One builder moved an Assistants MVP with little work. A team with its own history and retrieval found the port substantial. Another developer moved client work from Chat Completions to Responses for reasoning models.