On nine discovery-heavy CodeScaleBench tasks, Claude Sonnet 4.6 paired with the Sourcegraph MCP server outscored Fable 5 running on a local checkout, winning six of nine tasks. Sonnet 4.6 plus Sourcegraph MCP scored 0.698 at $1.02 per quality point versus Fable 5 at 0.568 and $1.83 per quality point, with the margin concentrated on tasks that require finding where a symbol is used across repositories, vulnerability tracing, and cross-organization dependency following.