Epoch AI found a brief 'first mover' price premium in one AI benchmark — does that pattern actually hold everywhere?

This explores whether the pricing pattern Epoch AI found is a general law of AI markets or a narrow observation. The short answer: the collection has one direct study, so it can't tell you how universal the effect is. Epoch AI tracked five benchmarks and found that the cost of reaching a given score drops about 66% per quarter right after that score first becomes state of the art. Within two years the drop slows to about 32% per quarter. Their explanation is a temporary monopoly: whoever reaches a new capability first can charge for it until competitors arrive Do AI performance prices fall faster when capabilities debut?. Five benchmarks is a pattern, not proof it holds everywhere. Nothing else in the collection tests the effect on other benchmarks.

The collection is more useful for a nearby question: why would a premium wear off so fast? One reason is that rivals are less different than they look. A study of more than 70 models answering 26,000 open-ended prompts found they often give strikingly similar answers, because they share training data and alignment methods Do different AI models actually produce diverse outputs?. If models end up in the same place, being first is a short-lived advantage, and the monopoly window closes quickly. Two other routes cut prices without needing a new frontier model. One is better execution harnesses, the tooling wrapped around a model. These raise scores on existing models without retraining, and the same setup carried over to newer models unchanged Can execution harnesses lift model performance without retuning weights?. The other is compact models trained on execution-grounded work. These reach the low-cost end of the cost-performance trade-off and compete with much larger models Does model efficiency matter more than peak capability for real work?. Each is a way that a capability someone recently charged a premium for becomes cheap.

There's a deeper caveat. The premium is attached to benchmark scores, and the collection questions what those scores mean. An analysis of 960 real job workflows found that agents win abstract contests but fail long, multi-step professional tasks. The authors call this a flaw in how benchmarks are designed, not a lack of model capability Why do agent benchmarks not predict real economic value?. Some benchmark gains may also come from contaminated test data, meaning the model memorized answers, even when the training does produce real reasoning behavior Can genuine reasoning activation coexist with contaminated benchmarks?. So 'state of the art' may partly describe how a lab ranks on a test, not what customers actually get. That changes what the premium is a premium for.

The takeaway you may not have expected: the premium is probably most universal where benchmarks measure things competitors can copy easily, and least universal where they measure messy, long-horizon work, because nobody may have reached real state of the art there yet. That's a hypothesis the collection suggests, not one it tests. If you want to follow this thread, start with the Epoch AI note for the effect itself, then read the benchmark-to-GDP note to see why the scoreboard behind it deserves skepticism.