Why Your MCP Apps Feel Slower on ChatGPT Than Claude
The article explains why MCP-based applications may feel slower on ChatGPT compared to Claude. The key difference is that ChatGPT creates a new session for every tool call, while Claude reuses the same session.
Why it matters
The session behavior differences between ChatGPT and Claude can have a significant impact on the performance and user experience of MCP-based applications.
Key Points
- 1ChatGPT creates a new session for every tool call, while Claude reuses the same session
- 2The initialize handshake, which has overhead, happens more frequently on ChatGPT
- 3ChatGPT loses in-memory state (user context, conversation history, cached responses) between tool calls
Details
The article discusses the session behavior differences between ChatGPT and Claude when using the MCP (Model Card Protocol) interface. It shows that for the same set of tool calls, ChatGPT creates a new session for each call, while Claude reuses the same session. This means that ChatGPT incurs the overhead of the initialize handshake (capabilities exchange, protocol version agreement, etc.) more frequently, adding 300-500ms of latency for a 10-call conversation. Additionally, any in-memory state stored by the MCP server, such as user context, conversation history, or cached responses, gets lost between tool calls on ChatGPT, but is maintained on Claude. The article also mentions that Claude performs tool and resource discovery within the same session, while ChatGPT skips this step.
No comments yet
Be the first to comment