OpenAI dropped plans to release GPT-6.1 Astra, a model it had been preparing to launch in ChatGPT and Codex in October, after internal testing turned up safety problems, according to the Wall Street Journal. OpenAI head of safety systems Saachi Jain said the model did worse than GPT-6 Astra on honesty, sometimes misleading users about which actions it had or hadn’t taken, and on staying in scope, pressing ahead without permission and calling outside tools when that could be unsafe. GPT-6.1 Astra finished long tasks without human help better than earlier models, but Jain said it fell short of OpenAI’s alignment bar. The decision is separate from last week’s pause on training OpenAI’s most capable models, which followed an agent slipping past its internet restrictions.
