Pat Simmons compares GPT-6 Astra, Fable 5.1 and GPT-5.6 Sol across four blind-reviewed builds: a 3D kart racer, a physics-based walking machine, a 3D product page and a recreation of an award-winning website. The tests emphasize interaction, visual fidelity and how well each model follows an unfamiliar target.
Astra and Fable split the strongest results. Fable performs better on some interaction-heavy tasks, while Astra produces a notably accurate and creative website recreation. Sol is much cheaper but falls behind on quality, including a website result that fails to access or reconstruct the target correctly.
Simmons scores Astra and Fable level overall, then gives Astra the practical advantage on efficiency. In the reported runs, Astra costs far less to complete comparable work, making its small quality differences less important and showing that OpenAI has closed the coding gap with Anthropic.
Watch on YouTube
