Datadog Scout 1.0: Cheaper AI SRE Investigations, Run More Often

Datadog says its first in-house model matches GPT-6.1 Sol high and Claude Opus 5.5 high on production investigations at 3 to 8 times lower estimated inference cost, and the point is not the saving itself but what it buys: investigations run on far more alerts, earlier. Datadog’s launch post for Scout 1.0 describes a model trained only to investigate production issues, powering Bits Investigation, the company’s AI SRE. For on-call engineers, the claim is that an AI first responder becomes cheap enough to run on every alert instead of the worst ones.

What Happened

Scout 1.0 is built on GLM-5.3 Flash and post-trained with reinforcement learning in investigation simulations made from historical incidents and real alert telemetry. Datadog says the simulations contain no customer data, and that training rewards finding the cause and backing conclusions with evidence. The model comes in effort levels. By Datadog’s figures, Scout 1.0 medium matches GPT-6.1 Sol high and Claude Opus 5.5 high at 3 to 8 times lower cost, and performs comparably to Fable medium and Astra medium at 12 and 19 times lower cost. Scout 1.0 xhigh matches Fable xhigh at 4 times lower cost.

Datadog engineering teams and a small group of preview customers already use it on production alerts, and Bits Investigation customers can sign up for the preview before general availability.

Why It Matters

Each investigation means forming hypotheses, querying telemetry and judging whether the evidence supports a conclusion, and every step is another model call. Datadog’s argument is that lowering the cost of those calls gives teams room to run investigations on more alerts and earlier in an incident. Cheaper calls also leave budget to spend more computation on the hardest incidents, which is what the xhigh setting does.

The evaluation deserves a careful read. Most results come from AlertEval, Datadog’s own benchmark of real alerts, archived telemetry and engineer-written root causes. On the third-party SREGym benchmark, Datadog reports Scout 1.0 medium beats its base model and surpasses GLM-5.3 at more than 4 times lower cost. Datadog also notes the scores measure agreement with expert explanations on a 0 to 100 scale, not the percentage of incidents solved. The technical report has the training details.

What This Means for Builders

Scout is trained on the tools Bits Investigation uses, and Datadog says it improves those tools based on what Scout’s investigations reveal. Our reading is that the model’s advantage is tied to Datadog’s platform rather than portable to other observability stacks. Datadog’s stated next step is remediation: simulating more failures and moving Bits from first responder toward being the only responder during an incident. Stack Archive tracked the start of that direction in Datadog DASH 2026: When Your Monitoring Platform Became an Agentic Operating System.

Source: Datadog

Content created with AI assistance and reviewed for accuracy.