GPT-6 Sol leaks, screenshots and release rumors

On TapNow, the author gave GPT-6 Sol a story log and style reference images, and it handled the script, character and location design, and video prompts, with images generated via GPT Image 2.5 and MiniMax H3.

A Turkish user says the cheap accounts they bought do half-baked, error-prone work that only becomes usable after GPT-6 Sol corrections, and complains its image generation is now worse than Gemini's.

A user says GPT-6 Sol is slow and dull while Opus 5.5 is fast and good, and compares Codex and Claude generating images with Qwen Image 2.1.

A comparison post placing Opus 5.5 and GPT-6 Sol Max image generation results side by side.

Given the same sunset ocean prompt, GPT-6 Sol finished in 44 seconds for $0.07, far faster than Claude and Astra, and the author asks which result was best.

On the same prompt, Opus 5.5 cost $1.27 and took 9m 35s, GPT-6 Astra $0.63 and 5m 4s, GPT-6 Sol $0.08 and 50s, and Luna under a cent and 43s.

A Japanese user says switching their main model to GPT-6 Sol High and image generation to GPT-6 Luna Light made quota drain much more slowly, suggesting Astra was the main consumer.

The author used GPT-6 Sol with Combos CLI to build a playable online 12-level card game, with in-game images generated by Sol calling GPT-image 2.5, and shared a prompt collection.

A full Japanese prompt for turning a full-body photo into a fisheye-lens subway car ad poster, which the author says should also work on 5.6 Sol.

A hands-on test generated one humanoid character each with Opus 5.5 and GPT-6 Sol: Opus 5.5 handled fingers and hair more naturally, while Sol's design was cuter but its hairstyle still fell short.

The author had Opus 5.5 swap GPT-5.6 in the original image for GPT-6 series figures, producing a chart comparing Sol and Luna cost-performance with Opus 5.5's performance.

The author says the GPT-6 models, especially Luna, are cheap and fun for image creation, and shares comparison examples of Sol, Opus 5.5 and Luna.