OpenAI Engineer Endorses Astra Over Flawed Evals

Ted Sanders, a research engineer at OpenAI, wrote that model evaluations frequently prove sloppy and fail to capture what matters in actual use. He gave the example of asking clarifying questions, which can lower scores even when that behavior proves helpful in real interactions. Sanders stated that Astra stands apart as the real deal. The model remains imperfect, he noted, yet it has crossed his personal trust threshold in a way no earlier model achieved. His comments appear in a single public post and reflect his direct assessment rather than any broader announcement from OpenAI.