What Happens When the Cost of Intelligence Drops 100x

That basic claim landed. The useful pushback was not "the curve is fake" but "cheap intelligence is not the whole story." Several people said the near-term bottleneck is latency. Even when budget is irrelevant, waiting on long model turns breaks interactive work. Others said reliability is the real dividing line. For many tasks, smaller and cheaper models are already good enough, so the contest is no longer "can the top model solve it" but "can a cheaper one do it consistently enough that you stop babysitting it." The strongest framing was that falling token prices change behavior more than they cut bills. Some expect a Jevons-style effect where cheaper inference causes far more inference. Agents will call other agents, contexts will get longer, and teams will rerun the same task multiple times for consensus or outlier detection. But commenters also sharpened the analogy. Jevons paradox is not automatic. It depends on demand elasticity, and some domains saturate. Lighting got cheaper and usage rose, but not enough to erase the efficiency gain. The open question is which parts of the economy have effectively unbounded demand for machine reasoning and which hit a ceiling fast. A second recurring point was commoditization. Open and Chinese models were described as already good enough for a large share of practical work at a fraction of frontier prices. That does not mean frontier models stop improving. It means the premium segment gets squeezed as easier tasks collapse into commodity inference. That fed a darker economic read from some commenters who think current prices are still distorted by VC funding, debt, or strategic subsidies, so today's cost curve may not map cleanly to a stable long-run business model. The more grounded takeaway was operational. As generation gets cheap, failure gets cheap too. That pushes systems toward high-volume experimentation, wider autonomous scope, and less human checking per step. In that world the important infrastructure is not just model quality. It is review, rollback, sandboxing, and ways to inspect and recover from bad outputs fast.