The San Francisco-based firm’s o1 system can go through a series of internal steps before responding to a user’s query or performing tasks like producing code or solving a crossword. (The Logic)

Talking point: An AI model that can reason through complex tasks could open up significantly more applications for the technology beyond answering questions. For example, AI “agents” can  perform chores that require multiple steps in different places, like scheduling an appointment. OpenAI is selling o1—internally code-named “Strawberry”—to workers in technical fields, like physicists that want help writing equations for quantum optics. The new model matched human experts and outperformed GPT-4o, the firm’s flagship language product, on maths and science tests. But OpenAI disclosed that o1 is also a bit better than GPT-4o at helping experts create biological threats; AI safety advocates like Yoshua Bengio have warned that bad actors could use advanced models to develop weapons.