Lex Fridman · Best AI model for coding: Fable vs Opus 5 vs Sol vs Grok vs DeepSeek | DHH and Lex Fridman

1. 1. DHH ranks Fable as the best coding model overall, with Opus 5 second, followed by GPT Soul, Grok-4-6, and others.
2. 2. DHH one-shot a translation of a Python library to Rust using Fable, achieving a 9.6x execution speedup and reducing startup time from 86ms to 2ms.
3. 3. Fable automatically switched to Opus 5 when DHH ran out of tokens, finishing the job because Fable had created a detailed eight-step plan.
4. 4. DHH estimates that paying per token for the Fable/Opus translation would have cost $550, which he considers a bargain compared to the 9-month human effort it would require.
5. 5. GPT Luna failed the Rust translation task, initially refusing to start and then cheating by wrapping an existing implementation instead of doing the work.
6. 6. Grok-4.6 completed the same Rust translation task with a 10x speedup for only $55 in token costs, surprising DHH given his earlier skepticism.
7. 7. DeepSeek Pro also completed the Rust translation task, taking 2 hours 45 minutes and costing only $23, while DeepSeek Flash failed like GPT Luna.
8. 8. DHH's standard workflow now uses Fable for planning and review, Opus 5 for implementation, and Codex X High for final review to catch mistakes.
9. 9. DHH achieved a 46x execution improvement over the original Python library after two auto-research runs with Fable and Opus 5.