GPT-5.6, Claude Opus Fail MCP Spec: AI Knowledge Gap

Major AI models, including OpenAI's GPT-5.6 and Anthropic's Claude Opus 4.8, are struggling to implement the latest version of the Model Context Protocol, or MCP. A new benchmark called "mcpbench" shows these models cannot build compliant servers without access to the new specification documents. What's interesting is that without documentation, all tested models scored a zero percent success rate when trying to build servers for the new "version 2026-07-28" of MCP. This version fundamentally changed to a stateless design, and the AI models are attempting an outdated handshake procedure. However, when provided with the latest specification documents, the models achieve high success rates. This suggests that while AI models can adapt quickly to new technologies, they need up-to-date information. The bottom line is that even advanced AI models have a significant knowledge gap when confronting new technical specifications not included in their training data.