probability: 87%

rationale:

Evidence suggests OpenAI currently possesses the internal capability to meet this milestone. In a TIME profile published 2026-08-26, Chief Scientist Jakub Pachocki stated that OpenAI has "already met its internal benchmark for an automated AI research intern" time.com. He described the upcoming Astra model as capable of implementing an experimental idea within OpenAI's codebase, running experiments, and returning results, or completing work that previously occupied a human researcher for a week time.com. This is functionally identical to the milestone described in the question.

The critical threshold is not internal capability, but whether OpenAI formally announces or releases the system using this specific framing. On a lenient reading, Pachocki's on-the-record statements in TIME, coupled with CEO Sam Altman's reported similar remarks gate.com, might already satisfy the requirement. Even under a strict interpretation requiring a formal product launch or official blog post, there is a substantial seven-month runway before the March 2027 deadline. OpenAI's DevDay, scheduled for 2026-09-29, presents a natural and highly probable venue for such an announcement.

The primary timing constraint is the recent safety slowdown. On 2026-08-07, OpenAI disclosed it could not rule out "Critical" cyber capabilities for Astra, prompting a pause on internal activities that did not meet hardened security requirements openai.com. By 2026-08-18, the company confirmed a two-week reinforcement-learning pause on deployment-bound models and placed its largest planned frontier RL run on hold openai.com. However, this pause predominantly affects the next frontier training run rather than the already-trained Astra model thezvi.substack.com. The company maintains its intention to ship Astra widely once new safeguards are cleared time.com, and market expectations heavily favor a release well before the deadline 2 sources.

The remaining residual risk stems from definitional strictness and potential tail-risk delays. A stringent resolver might reject magazine interviews as constituting an announcement by "OpenAI itself," requiring the company's official launch materials to use the exact "research intern" phrasing forbes.com. Under regulatory or political pressure, OpenAI could also opt to market Astra purely as a general agentic model and keep the "autonomous researcher" narrative strictly internal. Additionally, a further security incident or prolonged difficulties in mitigating Astra's cyber risks could theoretically push the release beyond March 2027. Nevertheless, weighing this narrow, near-term product milestone against other labs' much stricter quantitative thresholds for automated research affirms my high confidence in a timely release. Given OpenAI's active, voluntary use of the exact terminology in the press 2 sources and the long timeframe, a qualifying announcement is highly likely, supporting an 87% probability.