GPT-6 Sol leaks, screenshots and release rumors

Sol cuts factual errors from 8.4% to 4.5% at half the price but only moves the intelligence index from 47 to 48, and Luna is similarly cheaper with flat capability, the author says.

A Japanese user says GPT-6 Sol feels close to Astra in practice while using about 80% fewer tokens, making it strong value.

A parody written in the voice of The Godfather jokes about Sol and Opus 5.5 launching on the same day and Opus 6 already being on the way.

A Spanish-speaking user says heavy use of Codex and Claude still left 70-80% of the weekly limit and gives both a perfect score.

A video shows GPT-6 Sol on Combos Fun iterating faster with fewer tokens, and the platform offers a Sol prompt collection to help users start building games.

The poster says GPT-6 Sol costs half as much as 5.6, but the usable amount in Codex doesn't feel twice as large.

A hands-on comparison finds Astra is still stronger, but GPT-6 Sol is a clear upgrade over GPT-5.6 Sol.

A screenshot post lists API prices, cache discounts and AutomationBench scores for Sol and Luna, saying Sol beats Opus 5 at about 9% of the cost per task while Astra remains the flagship.

After testing, the poster says GPT-6 Sol is cheaper but clearly weaker and Opus 5.5 is no leap in intelligence, concluding the closed labs are stalling.

Microsoft researchers propose a dual-graph memory architecture that, combining GPT-6 Sol, Luna and Claude Opus 5.5, reaches a 53.08 RACE score on DeepResearch Bench.

Given the same prompt to build an interactive planet website, Sol produced 3 planets in 10 minutes using 1% of its quota, while Opus took 26 minutes to build a full solar system using 16%.

A Japanese blogger who tried GPT-6 Sol and Luna describes them as a half-price version of 5.6 Sol, and says they are bullish on AI downstream applications and investment.