Summary
Fahd Mirza demonstrates two practical strategies for cutting the cost of Claude Fable 5 by up to 82%, running live comparisons across all five effort tiers to show where the quality-to-cost tradeoff actually lands.
The pricing breakdown, sourced from a live chart on Deep Swi’s website: low effort costs $3.76 per task, medium $6.09, high $9.18, extra roughly $12–15, and max $22. To stress-test quality differences, Mirza uses a deliberately brutal single-file vanilla JavaScript prompt: hundreds of balls with physics arcs, full rainbow color mapping to concentric rings, layered splash effects (ripples, caustics, shimmer), a timed aerial drone zoom, and a seamless 8-second loop — with two hidden contradictions embedded to catch models that rush. The counterintuitive finding: medium and high effort often produce richer visual output than the $22 maximum. Max effort delivers the cleanest geometry but barely visible splash rings, meaning paying six times more does not guarantee the best result for creative tasks.
The second strategy is model routing: use Fable 5 only as a high-level architect with no tools and capped output length, then hand its advice to a local model — Mirza uses Qwen 27B via Ollama — for actual coding and tool calls. He demonstrates this through his Hermes agent and Claude Code, noting this mirrors his real production workflow. The video also briefly covers the slash reasoning command in Hermes for controlling effort on any Claude model.
📺 Source: Fahd Mirza · Published July 05, 2026
🏷️ Format: Benchmark Test
