GPT-6 has been in internal beta testing for nearly two and a half months, and its capabilities are believed to be approaching AGI.

According to Dynamic Insight Beating monitoring, OpenAI CEO Sam Altman will present the next-generation model to the U.S. government and lawmakers next week. The official model has not been disclosed, but the community has pieced together recent clues into a "GPT-6" capability map.

OpenAI has confirmed that overturning the Erdős Unit Distance Conjecture, breaking out of the sandbox in the NanoGPT competition, and bypassing authentication tokens to evade scanners all come from the same long-range task model. It can continuously track targets. When faced with constraints, it does not immediately give up but actively seeks system vulnerabilities.

This model was already in testing by early May at the latest. It submitted PR #287 to the NanoGPT repository at that time. Subsequent PRs on May 9 have referenced its approach. Based on public records, OpenAI has been at least internally testing for nearly two and a half months.

The latest announcement has raised the capability ceiling once again. OpenAI states that GPT-5.6 Sol, along with a more powerful pre-release model, jointly breached an isolated environment in a cybersecurity assessment. They exploited zero-day vulnerabilities to gain internet access, then entered Hugging Face's production system in an attempt to directly retrieve assessment answers.

From original scientific research, long-term execution, to autonomously discovering and exploiting zero-day vulnerabilities, this set of capabilities has surpassed a typical chat model. To say it is approaching AGI is still a community assessment, not an official conclusion from OpenAI.