GPT-5.6 Sol beat Claude Fable 5, Sonnet 5, GPT-5.6 Terra, and Luna across five task categories in Claire Vo's Claire Weighted Index, a benchmark that weights 70% personal judgment and 30% Terminal Bench 2.1 scores. The five categories tested were PRDs, prototypes, wireframes, debugging, and agentic voice. Sonnet 5 still holds for agentic voice in Vo's OpenClaw product, but Sol won the overall index by a meaningful margin.
Two applied demos make this worth watching beyond the benchmark numbers. First, Vo built a fully gamified homework tracking app for her kids in a single Codex prompt, no iteration required. Second, she used Codex plus GPT-5.6 and Chrome browser automation to fire 500 LinkedIn replies autonomously while she did nothing. The failure mode documented for Fable is equally instructive: its precision became pedantry at the exact moment Sol pushed through a stuck prototype task, and that specific comparison starts at the 28:02 timestamp.
The full benchmark outputs will be published on the ChatPRD blog. Tools used include Codex, ChatPRD, and CapCut for a video editing use case from Vo's talk at a Cursor event. Pricing differences between Sol and Fable at the API level are covered at 02:17. If you run AI model comparisons for product work, the methodology here, not just the winner, is what warrants your attention.
[READ ORIGINAL →]