Jean-Philippe Bempel @jpbempel.bsky.social

Datadog released Scout 1.0, its first model built to investigate production incidents. Scout is built on GLM-5.3 Flash. Datadog post-trained it with reinforcement learning in simulated investigations made from real historical incidents and telemetry. Datadog scores it on AlertEval, its own internal benchmark annotated by human experts. Scores measure agreement with expert reference explanations on a 0 to 100 scale. At medium effort, Scout matches GPT-6.1 Sol high and Claude Opus 5.5 high at 3 to 8 times lower estimated inference cost. At xhigh effort it matches Claude Fable 5.1 xhigh at 4 times lower cost. On the third-party SREGym benchmark, Scout medium beats GLM-5.3 at more than 4 times lower cost. Scout is in a closed preview, and Datadog runs the AlertEval numbers itself. My read: Datadog says Scout medium performs comparably to Fable medium at 12 times lower cost, and to Astra medium at 19 times lower. Divide 19 by 12 and you get about 1.6. So on this benchmark Astra medium has about 1.6 times the estimated cost of Fable medium for a comparable score. datadoghq.com/blog/ai/introd…