I put Fable 5.2 in Claude Code head-to-head with GPT-6.2 Astra in Codex across 12 real use cases, including web design, slide decks, a Canva painting, video editing, a 3D virtual summit, trip planning, financial models, and a browser game. Both ran on high effort with the exact same /goal prompt, and I tracked the cost, tokens, and runtime of every run. Fable won 7 to 5 on output, but it cost 7.2x more and ran almost 5x longer, even though both models are priced the same per token.