GPT-6 Sol leaks, screenshots and release rumors

A user says they don't trust GPT-6 Sol and Luna right now, calling them clearly worse than their GPT-5.6 counterparts.

A poster says GPT-6 Sol drains their usage as fast as GPT-5.6 Sol and asks whether Sol was supposed to be the efficient one.

The author ran the same JavaScript animation prompt on Claude Opus 5.5 and GPT-6 Sol, and two videos show the gap between them.

On OpenAI's new MentalHealthBench, GPT-6 Astra leads with 57.3, ahead of Sol at 53.9 and Opus 5.5 at 52.4, with no Chinese models on the list.

Both models build a 3D strategy game from the same task, and a video shows how the finished results differ in a hands-on comparison of Sol.

A Japanese post says Plus gives Astra 5–45 uses per 5 hours, GPT-6 Sol 15–150 and Luna 350–3000, a gap of up to about 67x, and offers tips for saving quota.

A developer complains with a screenshot that GPT-6 Sol is effectively unusable, that only the expensive Astra is viable, and considers going back to Claude.

Given the same prompt, Opus 5.5, GPT-6 Sol, Luna and Grok 4.7 each generate the house from Up, with Luna fastest and cheapest at 10:52 and about $0.03.

The author argues that only Opus 5.5 medium, Fable 5.1 xhigh, GPT-6 Astra xhigh and GPT-6 Sol medium are needed, instead of maintaining a long list of 120 models.

AgentOS now offers GPT-6 Astra, Sol and Luna plus Claude models, selectable by connecting a provider account through Vault, with a $sAGENT token mention.

The author tested several models on creative writing, pitting GPT-6 Sol against Opus 5.5 and Grok-4.7, with the stealth model space-bunny-alpha clustering closest to gpt-6.

A user argues GPT-6 Sol feels more like a minor upgrade to 5.6 Terra, with lower cost but no longer a true Sol, and questions whether it is a step back.