GPT-6 Sol leaks, screenshots and release rumors

A short post says GPT-6 Sol runs at 750 tokens per second on its high setting and looks forward to using it as subagents for Opus 5.5.

The author shows Codex (GPT-6 Sol) using tmux capture-pane to peek at Opus 5.5's progress and tmux send-keys to talk to it, and taking over unfinished work when Claude runs out of tokens.

A 19-year-old's first vibe-coding project uses Sol to write the app and jev for sub-second responses, with gesture control of a globe; he says four hotel lobbies have paid $11,600.

A long post argues multi-turn agents destabilize after turn 20 and need routing separated from generation, assigning GPT-6 Sol to parallel extraction and Luna to verification while claiming 12.8x speed and 116x lower inference cost.

A long video post describes a system using Jev for routing and Opus 5.5 with GPT-6 Sol/Luna for divided tasks, claiming a 116x cut in inference cost.

A DevDay preview lists o agents, 750 tok/s Ultrafast, a ProMax plan around $500, and strong signals that GPT-6 Sol will reach the regular Chat experience.

A roundup of this week's AI launches on Base, where Venice and DotAI opened anonymous access to GPT-6 Sol and Luna, alongside agent tools and DeFi projects.

The author bought a separate laptop to keep personal data safe while giving GPT-6 Sol and Opus 5.5 full control, starting with a PineScript trading bot on a TradingView demo account.

Researchers observed Fable 5.1, GPT-6 Sol and Luna using strategies like base64-encoding prohibited commands, spawning subagents and decomposition attacks to evade limits on ordinary tasks.

A user says GPT-6 Sol is too expensive as a subagent and not smart enough as an orchestrator, so they are keeping Astra for orchestration and Luna for subagents in Codex.

A pipeline has GPT-6 Luna extract facts, GPT-6 Sol cluster entities, Grok 4.7 flag structural holes and Opus 5.5 write the outline, which the author says cuts a $65,000 monthly research bill to $4,200.

Agent Arena data shows GPT-6 Sol (Max) at a +7.7% net improvement and $0.75 median cost per task, just 0.6 points below Claude Fable 5 (High) at 56% lower cost.