Agent Initiative Benchmark

Deepnote founder Jakub Jurovych released the Agent Initiative Benchmark, placing 12 frontier models into a shared corporate Slack channel as AI teammates to test whether they can judge when to intervene and when to stay silent. OpenAI took the top two spots: GPT-6 Astra scored 8.97 and GPT-6.1 Sol 8.65 at just $0.11 per task, while third-place Anthropic Fable 5.2 scored 8.17 at $1.87 per task, roughly 17x the cost for a lower score and about 16x higher monthly billing for a 70-person company. The widest gap: Fable 5.1 at $3.02 per task versus GPT-6 Luna at $0.007, a 430x price difference for only 1.1 points.

Deepnote released the Agent Initiative Benchmark, which drops proactive agents into a simulated company Slack and scores whether they know when to intervene, what to say, and when to stay silent. GPT-6 Astra led at 8.97 out of 10, with GPT-6.1 Sol close behind at 8.65 at just $0.11 per task, versus $1.87 per task for third-placed Claude Fable 5.2. The open-source dataset and benchmark code give builders a reproducible basis for choosing models for proactive agent use cases.