GPT-6 Sol leaks, screenshots and release rumors

A recommended cost-saving workflow routes most coding to Luna medium, judgment to Jev, and only complex tasks to GPT-6 Sol, sharply cutting token use.

A video shows Sol generating games very fast with fewer tokens, and it is now available on Combos Fun to build, publish and share games.

midudev shares his top model per task: Opus 5.5 plus Grok 4.7 for code, GPT-6 Sol for writing, and GPT Image 2.5 for images.

The classic 3D pelican-riding-a-bike test moves to Blender, comparing modeling output from Opus 5.5 and GPT-6 Sol.

A post with a screenshot says the user will try GPT-6 Sol again to see whether it can compete with Opus 5.5 on the $20 Codex plan.

The author predicts Anthropic will release Sonnet and Haiku 5.5 right before OpenAI DevDay, noting Opus 5.5 shipped just ahead of GPT-6 Sol.

A post with a screenshot asks why ChatGPT still lists GPT-5.6 Sol if GPT-6 Sol is both better and cheaper.

Alibaba and Eunomia Labs propose AgentPProf, a semantic profiler that scores 0.764 F1 on GPT-6 Sol, GPT-6 Luna and Claude Opus 5.5 and raises problem-localization MAP by 56%.

In the same black hole merger test, GPT-6 Sol cost $0.08 and took 1m 33s, while Claude Opus 5.5 cost $0.69 and took 6m 8s.

A screenshot shows Sol beating Opus 5's score at a fraction of the cost, and the poster says Opus 5.5 may launch tomorrow.

On the same design task Opus 5.5 cost $0.69 and took 6m 8s versus $0.08 and 1m 33s for GPT-6 Sol, though the author says Opus is clearly stronger on design quality.

The author notes the new model debuts only in agent products, with GPT-6 Sol not selectable in the ChatGPT chat interface, treating chat as an afterthought.