New AI skills launch at a premium on closed models — so their prices look like they're crashing, but it's really just monopoly pricing fading fast.

This explores why the cost of a fixed level of AI capability seems to fall faster on closed, frontier-lab models than on open-weight ones. The corpus doesn't directly compare closed and open price curves. It does contain a mechanism that explains the pattern, and that mechanism shifts the question: the speed of a price drop mostly reflects how high the price started, not whether the weights are open.

The key finding is that AI performance prices fall fastest right after a capability first becomes state of the art, by about 66% per quarter. Within two years the decline slows to about 32% per quarter Do AI performance prices fall faster when capabilities debut?. The explanation offered is temporary monopoly pricing. Whoever reaches a new capability first can charge a premium until competitors catch up. New capabilities almost always debut on closed models, so closed models are the ones that start with inflated prices and then fall steeply. Open models usually reach a capability months later, after the monopoly premium has already been competed away. Their prices start lower and have less room to fall.

The size of that delay matters. Open-weight models trail the closed frontier by about four months, and that gap has recently grown from about three months How far behind closed models are open-weight models?. Four months is long enough for the steepest part of the price drop to happen before an open model offers the same capability. Something else is happening at the same time, though. Open models small enough to run on a consumer GPU are improving faster than frontier models, at about +125 versus +80 LM Arena Elo per year Are open models closing on frontier AI faster than expected?. So open models are closing the capability gap fast, but they compete on price from the beginning instead of following a launch-premium curve down.

Usage data shows the market effect. On OpenRouter, open-weight models such as DeepSeek and Qwen serve over 30% of tokens even where closed models lead on capability, because users weigh cost and latency against raw performance How much do open-weight models actually get used in practice?. Cheap open alternatives are likely part of what erodes closed-model premiums. Each good-enough open release puts a ceiling on what a closed provider can charge for that tier of capability. A further pressure on price comes from inference: smaller models given more compute at answer time can match larger ones on hard prompts Can inference compute replace scaling up model size?. A frontier-sized model is therefore not the only way to reach a given performance level.

In short, the corpus doesn't show that closed models get cheaper because of something about being closed. They fall faster because they launch at monopoly prices that competition then brings down, and open models are a big part of that competition. If you want to verify the premise yourself, the corpus doesn't have a direct side-by-side of closed and open price curves. The pricing note is the place to start.