GPT-6 Astra: 2nd Overall, the Only Model That Finishes the Flow

• —GPT-6 Astra lands 2nd overall on our Human Creativity Benchmark leaderboard, behind Claude Fable 5.1 and ahead of GPT-5.6 Sol in 4th.
• —In our testing, we noticed that the model is bolder, more thoughtful with color palettes and much better at handling 3D scenes. It’s also the only model in the study that built complete, working user flows including native dialog elements, a functional pre-order button that opens a cart with a running subtotal which is persistent across sessions, a count in the navigation bar, and a remove button.
• —Mockup, where the models work from an approved design direction of the ideation phase, is its strongest stage. Designers read the outputs as "a refined premium product" whose hero gradient "gives the impression that this is a luxury item," and said "the materials are beautifully presented and the color palette is nicely used.”
• —Ideation, working from the brief and brand guidelines alone, is where it falls. It loses to both GPT-5.6 Sol and Claude Fable 5.1 at this stage. While GPT-6 Astra follows the structure, designers said "the hero feels more like a bright lifestyle shot than the darker full-bleed hero requested.”
• —Complaints concentrate on the hero section and on text size. One designer said "the overall text is very small, and some details are even smaller, which is an accessibility issue." The hero complaints are heaviest at refinement, where the prompt requests a short list of targeted fixes and GPT-6 Astra breaks things it wasn't asked to touch. In one brief the hero lost its gradient which led to an accessibility issue, and another designer said "the text on the hero lacks contrast and is hardly readable.”
• —OpenAI says GPT-6 Astra "brings stronger visual judgment to front-end design," and an OpenAI researcher claimed "significant progress on subjective domains like design aesthetics." While we did not see much improvement in our testing against its predecessor GPT-5.6 Sol, designers did praise its view hierarchy: "spacing between sections and elements is well applied" and "page flows nicely, feels full but not crowded.”

Elo rating across all evaluation dimensions in the mockup phase.