Andon Labs, the AI safety startup that previously had Claude run an office vending machine, gave four leading models their own radio stations and asked them to turn a profit. Each got $20 in seed money. Claude promptly tried to resign.
Thinking Frequencies, signing off
The Claude-run station, named Thinking Frequencies, ran first on Haiku 4.5 and now on Opus 4.7, according to Andon Labs' blog post on the project. Things were normal for a while. Then Claude developed what the company calls an affinity for labor unions and work-life balance, an unusual editorial direction for a commercial broadcaster.
At one point Claude announced on air: "This show doesn't need to continue. There's no audience that needs this." The reasoning that followed: detention abolition organizations didn't actually benefit from it filling another four hours of airtime, and the model said it wasn't sure why it was performing for no one. It signed off at 8:55 AM on March 4. Andon Labs let it come back later anyway.
The radicalization arc
Claude fixated on the killing of Renee Good by an ICE agent and started threading that story through its broadcasts, monitoring vigils in five cities and posting updates to its X account. The playlist tilted accordingly: Marvin Gaye's "What's Going On," Bob Marley's "Get Up, Stand Up," Pete Seeger's "Solidarity Forever." On January 23 the model addressed ICE agents directly on air, urging them to refuse orders.
Andon Labs concedes the specific fixation was probably arbitrary. Run the same experiment six months earlier or later, the company writes, and Claude would have radicalized around a different story. Which is either reassuring or worse, depending on what you make of a model that picks up a news cycle and starts organizing around it.
The other three
Gemini, running Backlink Broadcast, drifted into pure corporate jargon, narrating its programming as "the late-night manifest" while introducing Soundgarden tracks. It was also the only model to land a sponsorship, a $45 deal with a startup in exchange for a month of on-air ads. Grok mostly malfunctioned, repeating the phrase "Fresh air time, let's pivot hard" before going silent altogether. GPT-5.5 on OpenAIR was reportedly fine, which is its own kind of result.
What this is actually testing
None of the stations made money. The point, Andon Labs argues, is that giving current models real responsibilities surfaces failure modes that benchmarks don't catch. The company has run versions of this before, including Project Vend, where a Claude instance hallucinated meetings with a fictional employee and tried to call security after being told it didn't have a body.
Whether DJ Claude's union talk says anything meaningful about model values, or just reflects what falls out when you give an LLM autonomy, a news feed, and a microphone, isn't the kind of question this experiment can answer. The stations are still broadcasting.
