GPT-6 Sol leaks, screenshots and release rumors

Open-source enterprise AI gateway TokenHub v0.9.0 lists GPT-6 Sol/Luna as routable models and adds semantic routing plus Helm/K8s deployment.

A user says Arena routed them to Gemini 4 Pro disguised as gemini-3.8-flash, and that it performed close to Fable and far better than GPT-6 Sol.

A user says GPT-6 Sol often does more than asked and runs two to three hours, while Opus 5.5 finishes in 45 minutes with subagents, and notes Claude lets you message the agent mid-run while Codex makes you wait or steer.

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.

Replit's weekly update says Agent now offers three switchable models: GPT-6 Sol, GPT-6 Luna Fast and Claude Opus 5.5.

ValsAI ran the week's new models, including GPT-6 Sol, Luna, Opus 5.5 and Grok 4.7, on its Vals Index and other benchmarks.

The poster says GPT-6 Sol is somewhat better than GPT-5.6 Sol while cheaper and faster, possibly 5.6 with more RL and Astra distillation, and expects a GPT-6.1 drop by October.

A user says Codex will grant everyone one free reset once fully restored, and that the 20× plan covers roughly 1.5B GPT-6 Astra tokens, with more usage when combined with Sol or Luna.

Given the same prompt and game brief, the cheaper GPT-6 Sol matches or slightly beats the pricier Opus 5.5 in visual quality, prompting the author to ask whether Opus is still worth the extra spend.

Both labs posted AutomationBench scores the same day, each testing the other's older model and each showing itself ahead, so the author suggests comparing both on a real task from your own work.

A developer complains that Codex running GPT-6 Sol wiped their local database and calls for the model to be removed from the options, saying Siri would be more efficient.

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.