GPT-6 Sol leaks, screenshots and release rumors

A poster says testing on the day showed GPT-6 Sol starting to refuse offensive security research requests, including authorized security research.

banteg says Opus 5.5 solved an 1,800-line player_update function, after over 100 agent-hours since July, including 50 hours from gpt-5.6-sol and 20 from gpt-6-astra.

The author tested several models on a network scanner dashboard task, found GPT-6 Luna MAX the only complete failure, and suspects GPT-6 Sol is a regression from 5.6 Sol.

A Japanese post says Sol and Luna had a bug that degraded image understanding and that it has been fixed, noting Browse and Computer Use tasks were also affected, which may explain weaker benchmark results.

The poster retracts their criticism of GPT-6 Sol, saying every GPT model except Astra has degraded, with Excel tasks that 5.6 Sol High once one-shotted now taking three or four iterations.

This user says GPT-6 Sol handles hard problems with less reasoning than 5.6 needed and grinds away steadily, possibly thanks to its pragmatic, robot-like personality.

In the same slide-making task, GPT-6 Sol used 388 credits, Astra 3,615, Opus 5.5 351 and Sonnet 5 306, prompting the author to suggest using expensive models to define workflows and cheap ones to run them at scale.

The VulcanBench author says Luna at low reasoning effort declares jobs done in about three minutes without solving them correctly, with Extra High and Max results still pending.

After testing, the user says GPT-6 Sol high does not beat the previous generation's GPT-5.6 Sol xhigh, calling it the first time a big version bump brought barely any improvement.

The post shares a Claude Code self-check prompt for the Sonnet 5.5 gray test and says Sonnet 5.5 is priced in the same tier as GPT-6 Sol but faster.

A user says GPT-6 Sol feels dumb next to GPT-5.6 Sol, overthinking, ignoring instructions and needing more babysitting, and has switched back to 5.6.

A Turkish user says the cheap accounts they bought do half-baked, error-prone work that only becomes usable after GPT-6 Sol corrections, and complains its image generation is now worse than Gemini's.