Uber's $1,500/month AI limit is a useful signal for AI tool pricing

The strongest throughline was that per-token prices are likely headed down for routine work, even if frontier labs try to raise effective prices in the short term. Commenters pointed to Chinese and open-weight models like DeepSeek, Qwen, Kimi, GLM, and MiMo as good-enough substitutes for many coding tasks, especially when paired with a harness that routes simpler work to cheaper models. That shifted the useful unit of analysis away from “price per token” and toward “cost per completed task,” because newer agents often burn far more tokens to get a somewhat better answer. Plenty of people said the obvious optimization is no longer “always use Opus” but “use the smallest model that clears the bar.” There was also a blunt economic argument underneath all of this. If token revenue falls like a commodity while data centers are financed like long-lived infrastructure, the AI labs face a nasty mismatch. Some commenters think that means current API prices are still too low to cover the full capital buildout. Others think inference is already much cheaper than frontier labs charge, and that subscriptions are the subsidized piece, not enterprise API pricing. Either way, few people believed today’s setup is stable. The practical expectation was a split market. Expensive frontier models for the hardest work, cheaper open or older models for most of the volume, and more companies pushing usage controls instead of letting engineers spray tokens into autonomous loops. On the engineering side, the comments were much more grounded than the usual “AI replaces coding” talk. Heavy users said large models are often overkill for day-to-day development. Flash-tier models are good enough for small changes, refactors, tests, and internal tooling if a human stays in the driver’s seat. The real leverage is in the harness, prompt discipline, hooks, skills, and model selection, not just buying the smartest model. A recurring complaint was that unconstrained agent workflows generate giant PRs, overcomplicate code, and make review worse. People who reported good results almost always described tighter process, smaller tasks, and deliberate human supervision. That led to a broader conclusion. Uber’s cap is not proof that AI coding is worthless, nor proof that it delivers huge ROI. It is proof that companies are now forced to ask boring grown-up questions about budget, routing, lock-in, data handling, and what work actually merits frontier pricing. That is a healthier place than the earlier phase where the winning strategy looked like burning as many subsidized tokens as possible.