I benchmarked Claude Code orchestration against a solo session. My configured workers ran 0 times.
retained text
I had a Claude Code setup with a main-thread orchestrator and three configured workers (Opus, Sonnet, and a read-only scout). I wanted to know whether that setup was actually helping, so I compared it with an Opus 5.5 session working alone. The main benchmark used a private \~35k-line Python project. I created two bundles of four real, reverted bug fixes, held back the relevant tests as hidden checks, and ran the full 1,263-test suite for regressions. Each setup got one run per bundle. **Across those two bundles:** * Fable 5.1 orchestrator: 8/8 tickets, 0 regressions, $12.84 estimated list-price cost, 20.4 minutes * Opus 5.5 orchestrator: 8/8 tickets, 0 regressions, $5.79, 21.6 minutes * Opus 5.5 solo: 8/8 tickets, 0 regressions, $6.15, 24.4 minutes The surprising part: **neither orchestrator used a configured worker.** That was true across all eight orchestrator sessions in my two small benchmarks, even though the four tickets in each main bundle touched separate files. A separate pilot repeat did call a built-in agent, but not one of the configured workers. Two blind Claude reviewers preferred the solo solutions slightly, although they also incorrectly marked down one Fable solution for including evidence that the project instructions actually required. So I wouldn't treat those review scores as definitive. This is a small case study: one main run per cell, and a pilot repeat of one cell cost 29% more than its main run. Since the configured workers were never used, these results **do not tell us whether delegation pays off when it actually happens**. They did convince me to make direct work the default in my own template and require a specific reason to delegate. I published the method, results, limitations, and transcript measurement script here if you want the details: [https://github.com/Dealwatch/claude-code-orchestration-benchmark](https://github.com/Dealwatch/claude-code-orchestration-benchmark)
document fetched 2026-09-25 00:23 · canonical https://reddit.com/r/ClaudeAI/comments/1wp119w/i_benchmarked_claude_code_orchestration_against_a