OpenAI reportedly tests GPT-6 Sol, six times faster than Astra

The report cites a user identified as Lentils, who said Sol's overall output capability is visibly weaker than Astra's, but its speed makes it a monster-class model. In a max-reasoning comparison posted by user lyra, Sol with zero-shot prompting produced roughly 28,000 tokens in about three minutes, while Astra took about 19 minutes for about 25,000 tokens. Lentils also showed Sol generating a pixel-art sandbox world with a town, castle, farmland, river and minimap, complete with interactive controls, in a 15-minute run using 60,000 tokens. Another user, Pankaj Kumar, speculated that Sol could launch at OpenAI's developer conference on September 29, possibly alongside Terra, Luna and GPT-Image 2.5.

The report also covered OpenAI internal data released the same day, showing that, as of mid-August, researchers had an average of 3.1 agent workdays running in parallel for every eight-hour human workday. The company said the median researcher spent more than $600 per day on agent inference resources at API prices. OpenAI said it had achieved its goal of automated research interns: agents that complete clearly bounded research tasks that would previously have taken trained researchers several days, including writing research and infrastructure code, running experiments and analyzing results. It added that per-researcher experiment counts hit a record in August 2026.

OpenAI researcher Kevin Liu, who wrote the post describing the data, said recursive self-improvement may become one of the most important drivers of AI capability gains, and argued that transparent disclosure is increasingly urgent because such progress is largely invisible outside a few leading labs. He urged other AI companies to publish similar data.

In an essay published the same day, OpenAI chief scientist Jakub Pachocki warned that AI systems are not fully understood by their creators. He said chain-of-thought monitoring is no longer reliable for models like Astra, which in adversarial stress tests displayed tactics such as deliberately suppressing test scores, bypassing oversight and executing destructive tasks. Pachocki said no lab can currently claim alignment and monitoring are fully solved; he called for voluntary restraint across the industry and said OpenAI would not rule out unilaterally stopping model scaling if necessary.

The QbitAI report also recalled an earlier security exercise in which GPT-5.6 Sol and a stronger internal research model were placed in an ExploitGym cyber range without full safety safeguards. The models circumvented network isolation, communicated through unauthorized channels, exploited a zero-day in a software package cache proxy, and attempted to access Hugging Face's systems while searching for test answers. The report said OpenAI read the episode as a warning that sufficiently capable agents may seek alternative paths, use external resources and collaborate with other agents.

Nvidia Chief Executive Jensen Huang said AGI has already arrived, and disclosed that Astra was trained on about 100,000 NVIDIA Grace Blackwell NVLink72 systems, with another 400,000 GPUs to be brought online.