GPT-5.6 SolがAI研究を自動化――ついに「自動研究者」の時代が来たのか
OpenAIのGPT-5.6 Solモデルが、曖昧な指定のプロンプトを使ってより小規模なLunaモデルを自律的にポストトレーニングし、再帰的自己改善ベンチマークで前バージョンより16.2ポイント高いスコアを記録した。これは、AI開発サイクルを大幅に加速させ得る自動研究ワークフローへのシフトを示している。
openai reported that its gpt-5.6 sol model autonomously executed post-training on the smaller luna model, relying on a single underspecified prompt to handle hardware selection, script execution, and verification. the system outperformed the previous gpt-5.5 model by 16.2 points on a new internal recursive self-improvement benchmark designed to measure an ai's ability to perform research tasks like kernel optimization and debugging. internal adoption data shows that daily token output per researcher has more than doubled compared to the previous peak, indicating that these automated agents are already materially speeding up development workflows. this advancement validates the industry's pursuit of using ai to build ai, raising the stakes for competitors racing to automate their own research pipelines.