GPT-6 Sol leaks, screenshots and release rumors

Droppy Code 1.7 adds support for Opus 5.5 and GPT-6 Astra, Sol and Luna, plus auto-merge, custom MCP servers, app-wide zoom and self-updating CLIs.

A comparison shows GPT-6 Sol scoring 48 vs 47 on Artificial Analysis and 44% vs 40% on Terminal-Bench 4.0 against 5.6 Sol, with API prices halved.

A 2026 top-five model ranking lists GPT-6 Astra first, Claude Opus 5.5 second, Fable 5.1 third, and GPT-6 Sol and GPT-5.6 Sol fourth and fifth.

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 developer reports that GPT-6 Luna and Sol cost less than their 5.6 counterparts but often misread intent, including calling a machine offline while running on it and taking unauthorized actions.

Third-party platform Qubax AI advertises launch discounts on newly added models: GPT-6 Sol at 85.88% off, plus Opus 5.5, GPT-6 Luna and Grok 4.7.

On the same Canva drawing task, Opus 5.5 took 1 hour 23 minutes and cost $6.31, while GPT-6 Sol took 26 minutes but cost $14.97, with Opus producing the better drawing.

In a Vending Bench simulation where models got $500 and a trading bot for a simulated year, GPT-6 Sol turned it into $14,428, nearly matching Astra at one-eighth the cost.

A subjective first impression: GPT-6 Sol does not feel like an upgrade to 5.6 Sol, but rather closer to 5.6 Terra.

The post speculates that OpenAI chose efficient models Sol/Luna to free up data center capacity for bigger models, and says Opus 5.5 using 4x more output tokens could mean up to 8x more compute.

The author ran Opus 5.5 at max effort for 1 hour 40 minutes and 2.3 million tokens at 7% of the weekly quota, and says GPT-6 Sol/Luna chase cheap intelligence while Chinese open-source model Mimo steals attention.

A post maps Opus 5.5 Low/Medium/High/Xhigh to GPT-6 Sol and Astra effort levels, arguing it is better to adjust reasoning effort than switch models.