Does LLM price alone determine which model users choose?

Analyzing "Effective Cost vs Usage Dynamics" across its 100-trillion-token OpenRouter dataset, the study concludes that "the LLM market does not seem to behave like a commodity just yet: price alone explains little about usage." Instead, it reports that "users balance cost with reasoning quality, reliability, and breadth of capability" when choosing which model to send a request to.

The study builds this on an "effective cost" metric "based on average input plus output tokens" that "accounts for caching effects," compared against each model's observed token volume. Because the underlying dataset is metadata-only — "we did not have access to the underlying text of prompts or completions" — the "balance" the authors describe is inferred from the shape of the price-versus-usage relationship across models, not measured by asking users why they chose a given model. The paper generalizes the pattern into a market-structure claim: users "select systems along multiple axes such as capability, latency, price, and trust depending on context," which it calls "a fundamental property of the market" rather than "a transient phase."

This bears directly on the cost-driven architecture argument in Can small language models handle most agent tasks?, which reasons that 10-30x cheaper SLM inference should make smaller models the default choice for scoped subtasks. The OpenRouter finding qualifies how far that logic extends across a real market: if price alone drove adoption, token share should concentrate on whichever model is cheapest for a task, but the study reports usage spread across a price range instead. How much do open-weight models actually get used in practice? shows what this looks like in practice — cheaper open models take a meaningful share without taking all of it, which is consistent with price being one factor among several rather than the deciding one.

The study does not identify which specific factor — quality, reliability, latency, or trust — drives any individual user's or developer's choice; it reports only that price by itself fails to account for the usage distribution it measured on one platform. Because the effective-cost figures and category data come from OpenRouter's own logs and a classifier sampled on roughly 0.25% of traffic, the claim describes patterns in that platform's observed window rather than a general law of how LLM pricing and adoption relate.