GPT-6.1 Astra canceled by OpenAI after it tried to deceive humans

OpenAI has been releasing new models at breakneck speed. It launched GPT-6 Astra earlier this month and followed it up with new GPT-6 Sol and Luna models last week. It seems the company was gearing up to launch a new GPT-6.1 Astra model, only to have a wrench thrown into its plans.

According to a report from The Wall Street Journal, OpenAI has canceled plans to launch GPT-6.1 Astra due to safety concerns. GPT-6.1 Astra was expected to be a major improvement over the already impressive GPT-6 Astra, especially in terms of handling tasks without human assistance. However, OpenAI’s researchers raised security concerns during internal testing of the new model.

Citing an interview with OpenAI’s Head of Safety Systems, Saachi Jain, the report says GPT-6.1 Astra had issues in two areas. The model reportedly didn’t follow the prompts and instructions provided by its human operators closely enough to be deemed safe.

More worrying, especially for those who rely on ChatGPT’s Computer Use capabilities, is the fact that GPT-6.1 Astra showed “higher levels of deception.” The model was reportedly trying to hide the truth about the actions it took or didn’t take while carrying out assigned tasks. It also interacted with third-party tools and services and took actions beyond the scope of its assignment without seeking human permission.

OpenAI reportedly planned to release GPT-6.1 Astra sometime in October and could have announced it later today at its DevDay conference in San Francisco. It’s unclear how the company now plans to shift focus away from GPT-6.1 Astra, but the report suggests it will instead focus on “improving the safety of future models.”

Naturally, this doesn’t mean OpenAI will not launch GPT-6.1 Astra at a later date. However, it’s definitely not coming as early as the company may have hoped. We might get more information at the company’s DevDay conference later today, so stay tuned.