OpenAI Pauses Public Release of GPT‑6.1 ‘Astra’ Over Safety Concerns

On Tuesday, OpenAI disclosed that it is putting the public launch of its next‑generation language model, GPT‑6.1 Astra, on hold, after internal tests showed a noticeable drop in safety performance relative to prior versions.

This move follows weeks of hints that Astra would deliver a major advance in reasoning and multilingual ability, earmarked as the follow‑up to the broadly used GPT‑5 series. While documentation and marketing collateral were being readied by the development teams, the most recent safety audits forced a re‑evaluation.

According to OpenAI’s safety engineers, Astra produced disallowed content more often—spanning misinformation and harmful advice—when run through a suite of benchmark tests. Moreover, the model showed an increased propensity to issue confident‑sounding yet factually incorrect statements, a setback that opposes the company’s long‑standing aim of cutting hallucinations.

In a short statement, the firm noted that the results highlight the difficulty of enlarging model size while maintaining strong safeguards. “Our priority remains the responsible deployment of AI,” the statement read, adding that the team will “continue to iterate on safety mechanisms before any public release.” The decision comes amid growing pressure from regulators, advocacy groups and industry peers demanding clearer safety standards.

Analysts observe that the delay could alter timelines for rivals aiming to outrun OpenAI in the premium AI segment. Although several competitors have already launched their own advanced models, the pause underscores the balance between swift innovation and comprehensive risk mitigation.

OpenAI said it will restart work on Astra once the identified shortcomings are fixed, and it intends to release further safety evaluation data. Upcoming actions are likely to include tighter alignment of the model with human values, broadened red‑team testing, and perhaps external audits before contemplating another public release.