Raindrop's new agent debugger is 1,600x cheaper than GPT-5.6

If you have ever shipped an AI agent to production, you know the classic nightmare: the model calls an external tool three times, watches it time out every single turn, and then calmly tells the customer that everything went through successfully.

Catching those subtle, multi-step failures has always been a massive headache. If you send every trace through a frontier reasoning model, your inference bill will instantly outpace your revenue. If you use a cheap lightweight classifier, it chokes on long contexts and complex agent tool logs.

Raindrop just rolled out Signals 2.0, powered by an internal model pipeline called rd-signal-2. It builds task-specific binary classifiers designed to catch agent failures directly inside high-volume production traces.

According to Raindrop, rd-signal-2 approaches the accuracy of GPT-5.6 Sol xhigh at 1/1,600th of the cost, while beating GPT-5.6 Luna xhigh by 260x on price. The median trace classification takes just 100 milliseconds, and the system is already chewing through over 20 billion traces every month.

Reasoning at build time, code at runtime

Raindrop pulls this off by moving the expensive reasoning from runtime to build time.

Instead of paying a frontier model to read hundreds of thousands of tokens on every single user interaction, rd-signal-2 analyzes historical traces upfront and writes deterministic code to extract the exact evidence needed.

Take the repeated failure example: the generated code first checks whether the agent called the same tool three times with identical parameters and received errors. If that condition is not met, the trace is marked clean immediately without waking up an LLM.

Only the ambiguous candidate traces are forwarded to a specialized classification head and an in-house reasoning model, carrying only the tiny slice of relevant context. Raindrop also introduced Signal Builder with Zero Data Retention, sandbox isolation, and an MCP interface, opening the door for regulated sectors like healthcare.

Why it matters

We think Signals 2.0 solves one of the dirtiest open secrets in generative AI: running automated LLM-as-a-judge pipelines across millions of production agent traces has been economically impossible for most teams.

By turning behavioral detection into a two-stage filter—cheap deterministic code first, compact model reasoning second—Raindrop has made full-scale agent observability cheap enough to bundle into its core platform at no extra charge. If you are running autonomous agents in customer-facing workflows, being able to detect hallucinations and tool loops in 100 milliseconds without paying a second cloud tax makes this an absolute no-brainer to try.

The real battleground now shifts from model inference to behavior definition. In Raindrop's own tests, writing four slight variations of a single failure description changed match rates on 2,000 production traces from 0.9% to 4.6%, with 67% of matches rejected by another variant. The bottleneck is no longer how fast you can classify failures, but whether your team can clearly agree on what a failure actually looks like.