rd-signal-2: State-of-the-Art Classification Models at Production Scale

Raindrop has launched Signals 2.0, a new model pipeline based on rd-signal-2 that builds task-specific binary classifiers using production traces.

rd-signal-2 approaches the high accuracy level of GPT-5.6 Sol xhigh while being 1,600 times cheaper than that model and 260 times cheaper than GPT-5.6 Luna xhigh.

Failures in AI agents do not occur in a single turn but arise complexly through numerous turns, tool calls, and interactions between sub-agents. Applying large frontier models to all traces is inefficient in terms of cost and speed, while small models have limitations in grasping complex contexts.

rd-signal-2 addresses this problem through an automated research loop. It analyzes production traces for each behavior, directly writes code to collect important context, and then trains a model optimized for the specific task based on that.

Additionally, it released Signal Builder, a platform that allows training and hosting custom classifiers with Zero Data Retention. This enables powerful classification capabilities even in strict security environments such as the healthcare sector.