introducing rd-signal-2: a frontier classification model that is 1600x cheaper than GPT 5.6 Sol. free to try in Raindrop, and available via a new API for training/hosting custom classifiers with Zero Data Retention.

read more at: rd-signal-2 balances precision/recall at a price that is still 280x lower than GPT-5.6 Luna xhigh. it powers signals and real-time issue detection in raindrop, allowing our customers to detect and accurately track sparse behaviors.

most LLM observability providers require customers to sample a very small % of traffic (and pay for the judges!) but agent failures are often rare. it's common for a behavioral anomaly to be < 1% of traffic. sampling can completely hide the problem.

you can say "the agent saying it transferred to a human, but didn't call the tool" or "the user complaining that web search results were stale" @raindrop_ai signals will write the code, assemble the context, train a custom classification head for your task, and self-improve.

rd-signal-2 also powers @raindrop_ai's real-time issue detection. many emerging failures are detected in under 5 minutes. because they share the same foundation, any raindrop issue can be instantly converted to a signal, compared in experiments, and more.

rd-signal-2 already processes around 20 billion traces a month. members of the @raindrop_ai team: - invented fraud transformer models at robinhood - built the first recommender system at pinterest - have built agents that have millions of customers - and more dm me to join

stop paying for judges. starting using signals. try the agent observability platform trusted by vercel, speak, and the fortune 100. read more at

@raindrop_ai Damn your motion design is on point! Looks awesome

@raindrop_ai thank you!

@raindrop_ai sweeet, great job! adaptive online observability is basically table stakes at this point

@raindrop_ai we agree.

@raindrop_ai “Stop paying for judges. Start using signals.” :chefs_kiss:

@raindrop_ai Typical raindrop W so glad I told @orchid_hq to use you guys

@raindrop_ai very excited to try this out!

@raindrop_ai hey @alialobai1 @hegargarcia 👀

@raindrop_ai @alialobai1 @hegargarcia would love your feedback :)

@raindrop_ai huge

@raindrop_ai thanks bani! :)

@raindrop_ai woahhh

@raindrop_ai give it a try!

@raindrop_ai very cool

@raindrop_ai Very interesting 🤔🤔🤔

@raindrop_ai well shit that sure is a lot cheaper

@raindrop_ai indeed

@raindrop_ai this is huge

@raindrop_ai thank you - please give us any/all feedback :)

@raindrop_ai will do excited for this :)

@raindrop_ai interested how this is implemented, really cool

@raindrop_ai Very cool!

@raindrop_ai thanks jeff!

@raindrop_ai thats crazy but nobody uses frontier models for classifiers anyway?

@raindrop_ai yes, so then you get very bad precision/recall. you can also see a comparison to 5.6 Luna (which is a good classifier)

@raindrop_ai dang this is good

@raindrop_ai thanks friend!

@rsdgpt @raindrop_ai smart play!

@raindrop_ai very impressive!

@raindrop_ai Wait. This. FUCKING BANGS......

@raindrop_ai thank you! which part resonated the most ooc

@raindrop_ai @Teknium @NousResearch @yeahfortommy this seems promising

@raindrop_ai sick dude

@raindrop_ai whoa this wasn’t on my bingo card but so excited and happy for ben :)

@raindrop_ai oo this is cool

@raindrop_ai oh my god i have a very good use case for this if feasible

@raindrop_ai dm me!

@raindrop_ai oss?

@raindrop_ai not this time! we do have plenty of other open source work though.

@raindrop_ai ngl my main question is how did you create the video. great job - looks amazing.

@sjwhitmore @raindrop_ai Fire

@raindrop_ai Super cool! Thanks for releasing it publicly

@raindrop_ai Cool stuff! @raindrop_ai team!!

@raindrop_ai 1600x cheaper only matters if the error rate holds. A classifier two points worse costs more in human review time than it ever saves in tokens, at any price per call. What does the confusion matrix look like against Sol on the same eval set?

@raindrop_ai Whoa

@raindrop_ai agreed!

@raindrop_ai I see that your Triage boy is free to try, saw someone made a token with royalities to your github. So you can claim the trading fees from it, if you want. GL chad

@raindrop_ai cheap inference meta

@raindrop_ai 1600x cheaper? Didn't even know a price could go that low.

@raindrop_ai constraining the problem to binary classification helps re: cost a lot.

@raindrop_ai Makes sense to me.