Datadog's Scout 1.0 cuts AI inference costs for faster production incident response

Datadog has built Scout 1.0, a specialized artificial intelligence model designed to investigate production incidents before human engineers get involved. The company trained this system specifically for site reliability engineering work—the routine but labor-intensive task of examining alerts, checking telemetry data, and narrowing down what actually went wrong when systems fail. What sets Scout apart is not what it can do, but what it costs to do it.

The economics matter because incident investigation is inherently expensive to automate. Each investigation requires the model to make repeated calls, forming hypotheses, querying data, and testing whether the evidence supports a conclusion. When teams rely on large general-purpose AI models for this work, those repeated rounds of questioning quickly drive up compute bills. Datadog's answer was to build something smaller and more focused. Scout 1.0 medium matches the performance of GPT-6.1 Sol high and Claude Opus 5.5 high on the specific scenarios Datadog cares about, but at an estimated inference cost three to eight times lower. Against other specialized models like Fable medium and Astra medium, the cost advantage grows to twelve and nineteen times lower, respectively. With additional test-time scaling, Scout 1.0 xhigh reaches cost parity with Fable xhigh while spending four times less to get there.

That cost structure changes what becomes economically practical. When automated investigation is expensive, teams reserve it for exceptional cases. When it becomes cheap enough, it can become routine—part of how every alert gets handled, not just the ones that seem important. Datadog's longer-term vision is for its Bits Investigation tool, powered by Scout, to act as an always-on first responder for every alert, gathering evidence and identifying likely causes before an engineer steps in. For on-call staff managing frequent alerts or dealing with incidents outside normal working hours, that shift could meaningfully reduce manual triage burden.

Datadog built Scout 1.0 on GLM-5.3 Flash and then post-trained it using reinforcement learning in simulated investigations. The company created those simulations from historical incidents and telemetry data from real alerts, using known root causes to train the model while carefully excluding actual customer data from the training process. Datadog also built its own distributed reinforcement learning infrastructure using open-source tools, and trained Scout to use tools modeled on Bits Investigation itself—so the model learned to gather evidence, test competing explanations, and support conclusions with data.

To verify that these skills would transfer to real incidents, Datadog created an internal evaluation framework called AlertEval. Engineers curated thousands of evaluation points by working through actual incident-response workflows, combining real alerts, archived telemetry, and root-cause explanations written by engineers who had already solved those problems. Scout was also tested on SREGym, a third-party benchmark for site reliability engineering tasks. On that benchmark, Scout 1.0 medium improved significantly on the underlying GLM-5.3 Flash base model and surpassed the full GLM-5.3 at more than four times lower cost. Scout 1.0 high proved competitive with leading models on the same benchmark while maintaining at least double the cost savings against all competitors and ten times the savings against Astra- and Fable-class models.

Datadog's own engineering teams are already using Scout to investigate production alerts, alongside a small group of customers in a closed preview. That internal use is deliberate and strategic. Datadog sells observability and monitoring software to engineering teams managing large, complex cloud systems. A model that can handle the first stage of incident investigation in Datadog's own operations becomes proof of concept for what it can do in customer environments. The company is testing Scout both in its own production systems and in external environments ahead of a wider release.

What emerges from this approach is a tightly integrated product strategy. Datadog is developing the model and the investigation product together, training Scout for the specific tools Bits uses and adjusting those tools based on what it learns from investigations. Scout 1.0 is not a standalone release dropped into an existing workflow; it is part of a controlled product ecosystem within Datadog's software stack. Preview customers are already using Bits Investigation powered by Scout on their production alerts, which means the real-world testing has begun.