GPT-6 Sol leaks, screenshots and release rumors

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.

A side-by-side video of the same prompt shows Opus 5.5 beating GPT-6 Sol on level of detail, a gap the author calls not even the same universe.

A Japanese post says Opus 5.5, GPT-6 Sol and Luna all arrived at once, and recommends a generative AI showcase in Hamamatsucho to hear hands-on impressions.

Augment Code says GPT-6 Sol and Luna are now available in Cosmos, and it will move its workflows off the GPT-5.6 series in the coming days while measuring the impact.

Dan Shipper says GPT-6 Sol and Opus 5.5 launched today and that the Every team has been testing both extensively on real work ahead of sharing conclusions.

A Japanese post says new models arrived on the day of a generative AI showcase, so the author is starting benchmark evaluations of Opus 5.5 and GPT-6 Sol/Luna.

The poster says GPT-6 Sol and Opus 5.5 are neck and neck on ErrataBench with a similar performance-cost curve, but both trail Astra and Opus 5.

A user running GPT-6 Sol at high with only 3 parallel tasks says it advanced just 3% in 14 minutes and expressed shock at the speed.

A Japanese user posts that GPT-6 Sol and Luna have launched with better performance at half the cost of GPT-5.6, plus one Codex reset.

Based on a chart, the author recommends Opus 5.5 at medium, GPT-6 Sol at high, and GPT-6 Luna at extra-high as preferred model tiers.

A free, open-source Claude Code plugin called fable-advisor v6.0.0 combines Opus 5.5, Fable 5.1, GPT-6 Sol and Luna into one AI team.

Scores for Opus 5.5, GPT-6 Sol and GPT-6 Luna appear on the Playground object-detection leaderboard, and the poster says Google has lost its lead in the category to three labs and four models.