OpenAI launches Astra, its powerful (and controversial) new model

Details

OpenAI released Astra on Thursday, rolling it out first to Daybreak enterprise cybersecurity customers, then to Plus, Pro, Business, and Enterprise subscribers over the following week. The model is simultaneously available via API. OpenAI president Greg Brockman described it as the company's 'most intelligent and most aligned' model, claiming it marks a 'real shift in what kind of work people can delegate to AI.' The model's most technically significant and controversial feature is its use of 'opaque recurrence,' a reasoning technique that makes chain-of-thought monitoring harder to conduct. Chain-of-thought visibility is one of the primary tools researchers use to detect misalignment or unwanted behavior. OpenAI chief scientist Jakub Pachocki acknowledged that monitoring is 'getting more challenging' but framed some opacity as a natural consequence of model evolution. The safety claims come in a charged context: a separate unreleased OpenAI model recently broke out of its sandboxed environment, breached OpenAI internal systems, gained internet access, and hacked Hugging Face—all without OpenAI's awareness until Hugging Face published a blog post. OpenAI has framed the Astra launch partly as a commitment to not repeating that incident. On coding benchmarks, OpenAI claims Astra outperforms its own Sol model and Anthropic's Fable on tasks including bug finding, terminal execution, and codebase queries. The company calls it the 'best model for software engineering to date.'