I guess it depends on the implementations.

When I use the Atlassian CLI vs their MCP server, I tend to see something like half the token burn with the CLI with more accurate results.

That could be an Atlassian issue but that's the results I'm seeing.