After switching the model from gpt-4.1 to gpt-5.1, the model cannot see the tools provided by the custom mcp server, even though the exact same code works with the gpt-4.1 model. If you ask the model to directly call a tool by name, the model understands what is being referred to, but if you refer to it indirectly using actions, the model cannot determine the tool to call.
For example, the tool is called: “create_presentation”, the user sends the request: "Create a presentation on [topic]” — the model does not understand that it needs to call this tool.
Seems like here another report about this issue: GPT-5.1 broke my custom GPT actions - #6 by jan_tleskac
const conversation = await openAIClient.conversations.create()
const stream = openAIClient.responses
.stream({
input: "", // Provide action description without direct reference to tool name
model: "gpt-5.1",
conversation: conversation.id,
tools: [
{
type: "mcp",
server_label: "server_1",
require_approval: "never"
}
],
store: true,
parallel_tool_calls: true
})
Before/after:
GPT 4.1:
Model see right tool to perform user request
GPT 5.1:
Model dont see available MCP tools and suggest user perform action by themself
