FABLE IS BACK! (And Sonnet 5 is here too)
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
A hands-on evaluation of Claude Sonnet 5, covering its benchmarks, real-world performance, and cost efficiency compared to Opus 4.8, GPT-5.5, and other models. Key findings: Sonnet 5 introduces Fable-like sub-agent orchestration capabilities to a mid-tier model, but it is inefficient — using far more tokens than competitors, running slowly, and often costing more than Opus in practice despite lower per-token pricing. The model's agentic behaviors (spawning sub-agents, breaking up tasks) are promising but undermined by insufficient intelligence, causing it to loop and waste tokens. Also covered: the lifting of US export controls on Claude Fable 5 and Mythos 5, benchmark contamination concerns (especially SWE-bench), and a leaked thinking trace showing verbose, unfocused internal reasoning. The author concludes Sonnet 5 is best suited as a sub-agent called by smarter orchestrators, not as a daily coding model.