GPT-6 Sol leaks, screenshots and release rumors

Researchers say GPT-6 Sol matches RetroChimera, the state-of-the-art retrosynthesis planning model spotlighted in Nature four days earlier, on their URSA benchmark accepted to NeurIPS 2026.

A user on the $20 plan ran Opus 5.5 for four hours and got a nearly production-ready full-stack app, and plans to have GPT-6 Sol review the code.

The poster says claims that GPT-6 Sol has gotten worse are wrong, and maintains it is stronger than 5.6 Sol.

After three days comparing GPT-6 Sol and Claude Opus 5.5, a user says Sol's usage limits are excellent but it missed too many details, and leans toward Opus 5.5.

Under the same limits (20 moves or 5 minutes), Opus 5.5 solved the cube in three moves in 1m 12s, while GPT-6 Sol made a wrong first move and timed out, and Jev and Laya failed to solve it.

Four models ran the same cube task with a 20-move or 5-minute limit; Opus 5.5 finished in 1m 12.87s with three moves, while GPT-6 Sol made a wrong first move and timed out after four.

A user praises GPT-5.6 Sol for its odd, cute ideas and publicly asks that GPT-6 Sol not replace it as the main model for Chat.

A short post asserts that Sonnet 5.5 beats GPT-6 Sol and says things are not looking good for OpenAI, with no evidence or screenshots.

A video puts games generated by Claude Opus 5.5 and GPT-6 Sol side by side and lets viewers decide which one did better.

A brief post only mentions cost efficiency with GPT-6 Sol, with no specific figures or screenshots.

A post with an image jokes that if OpenAI were confident in GPT-6 Sol, Sam would be drunk tweeting or Tibo would be taunting competitors.

The user says the quota Sol saves ends up spent on having Astra redo its work, and advises using Astra Light instead of Sol.