OpenAI 的新 Astra 數學突破——還是僅僅是一個巧妙的提示詞技巧？

an anthropic mathematician successfully replicated half of openai's claimed astra mathematical proofs using only the public fable model, generic prompting, and offline computation. this rapid partial replication suggests openai's reported breakthroughs may rely more on identifying problems amenable to specific search-and-verify techniques than on unique model capabilities. the scrutiny casts doubt on whether astra constitutes a genuine leap over existing models, especially given openai's opaque reporting on failed attempts. internal indecision at openai regarding whether to brand the release as gpt-5.7 or gpt-6 signals uncertainty about its competitive differentiation. the incident highlights the difficulty of assessing ai progress when companies present numerators without denominators, obscuring the true scope of model limitations.