Descriptions:
Nate Herk runs Claude Opus 5 and Claude Fable 5 through a battery of practical experiments inside Claude Code, tracking cost, time, and token consumption per task to give practitioners data beyond Anthropic’s official benchmarks. All tests use the same Claude Code harness, isolating model performance as the variable.
Results are nuanced and sometimes counterintuitive. On a large-codebase bug review task, Opus 5 came in cheaper ($4.22 vs. $5.30) but took slightly longer (13 vs. 11 minutes), with Codex judging Fable’s output marginally better on patch precision. For an Excalidraw diagram generation task with no verification loop available, Opus 5 produced a more organized layout while Fable’s was more visually dynamic — a judgment call with no clear winner. Most strikingly, on a LinkedIn post-plus-carousel content creation task, Opus 5 actually cost more than Fable ($8.22 vs. $6.17) and ran slower, revealing that Opus consumes significantly more tokens on certain task types, erasing its per-token price advantage.
Herk concludes that Opus 5 earns its place for verification-heavy coding work and complex knowledge tasks, but is not a universal drop-in replacement for Fable 5 — and may even be overkill for content creation where Sonnet 4.5 would suffice. The video serves as a practical guide for practitioners trying to optimize model allocation across different workflow types within Claude Code.
📺 Source: Nate Herk | AI Automation · Published July 24, 2026
🏷️ Format: Comparison
