This Hacker News discussion centers on the growing frustration among power users regarding **Claude Code’s** token-heavy "Fable" and "Opus" models. Users report that Claude is becoming increasingly verbose, inscrutable, and prone to "token-guzzling" sub-agent loops, which leads to premature exhaustion of weekly usage limits. A significant shift is occurring: many professional developers are migrating to **OpenAI’s Codex** and other alternatives, citing superior efficiency, more reliable performance, and more generous usage limits. While some users appreciate Claude’s sophisticated reasoning for complex architectural tasks, many feel the "vibe-coding" approach—where models make autonomous, unguided decisions—often results in unnecessary technical debt and wasted resources. Key themes include: * **Efficiency vs. Capability:** The debate over whether long-term success belongs to "smart" but slow, expensive models (Anthropic) or "fast, efficient, and consistent" systems (OpenAI/local models). * **The "Harness" Advantage:** Users are finding that better orchestration, local model experimentation, and human-in-the-loop workflows offer higher ROI than relying on "black box" agentic models. * **Provider Fatigue:** There is widespread exhaustion with fluctuating usage promotions and the perception that Anthropic’s models are being "nerfed" or optimized specifically to maximize token consumption.