Sonnet 4.6 with Sourcegraph MCP Server Beats Fable 5: 0.698 vs 0.568

Claude Sonnet 4.6 with the Sourcegraph MCP server scored 0.698 on nine CodeScaleBench tasks, beating Claude Fable 5, which scored 0.568. Sonnet worked without source code on disk and reached the codebase only through search tools. Fable 5 had a full local checkout of the repository and no retrieval at all. The cheaper model won six of the nine tasks and cost almost half as much per quality point.

CodeScaleBench is Sourcegraph's own benchmark for coding agents working in large, multi-repository codebases. For this comparison, the company picked nine tasks where the main job is finding the right code. Fable 5 is the first generally available model in Anthropic's Mythos class, a tier above Opus. At release, it posted state-of-the-art results on most public benchmarks, including Cursor's and Cognition's coding evaluations.

Search vs. reading from disk

The two configurations are opposites.

• Sonnet 4.6 + Sourcegraph MCP server. No source code on disk. The agent searches the code, resolves symbols, and follows references between repositories.
• Fable 5 with a local checkout. The entire repository sits next to the model, which reads files directly. There are no retrieval tools.

The final numbers:

• Sonnet 4.6 with Sourcegraph: average score 0.698, won 6 of 9 tasks.
• Fable 5 with local code: average score 0.568.
• Fable's cost per quality point is almost twice as high. Sonnet costs about half as much.

The cheaper model with a tool came out ahead on both quality and price. For teams that track what their agents cost, the second metric matters as much as the first.

Where exactly the gap comes from

Sonnet's advantage is not spread evenly across tasks. Most of it comes from tasks that require finding how a symbol is actually used across several repositories. Vulnerability tracing and dependency tracking across organizations lead the list. This is where reading a local copy directly helps little: the code you need may not be in it at all, or it may be too far away.

On simple tasks, there is almost no difference. Config tracing and a clean migration are handled well by both setups, and their results are close. Where the ceiling has already been reached, there is nothing left to win.

There is also a counterexample. On one of the tasks with cross-organization dependencies, the frontier model with a full local checkout did come out ahead. That detail speaks in favor of the methodology: the test was not fitted to a winner known in advance.

Sourcegraph has published a results explorer. For any task, it shows the runs of both setups side by side: the full prompt, the dialogue, and every tool call. You can see where one agent finds the right file and the other never gets to it.

Caveats you should not skip

First, this is nine tasks, not the whole of CodeScaleBench. Access to Fable 5 was restricted a few days after release, so the comparison includes only tasks with valid Fable runs completed before that cutoff. The author of the piece himself calls the result an early signal, not an established benchmark.

Second, the benchmark belongs to a company that sells the winning tool. That does not negate the openness of the runs, but independent replication on other tasks would not hurt.

Third, the test measures code navigation. Tasks that require deep reasoning inside a single file were left out. Sourcegraph names them as the next area to measure, and that is exactly where retrieval's advantage may disappear.

What to check when building an agent on a limited budget

The practical takeaway for teams building agents on real codebases: before upgrading the model, evaluate the quality of your code search. On cross-repository tasks, such as vulnerability search, dependency analysis, and change impact assessment, a cheaper model paired with good retrieval can outperform an expensive model without it.

Take 10–20 of your own typical tasks and split them into two classes: "find where this is used" and "understand one complex section." Run the first class on a mid-tier model plus MCP search, and the second on a flagship with local code. Compare cost per unit of quality, not just the final score. By that metric, Sonnet 4.6 with Sourcegraph beat Fable 5.