LLM API price survey: identical medians, but some models cost 33x more by provider

A pricing paradox in the LLM market

A comparison published on dev.to examined API pricing across 137 inference providers covering 4,033 models, then narrowed its focus to the 375 models offered by at least two providers where per-million-token input rates could be compared directly. The result is a market that looks simultaneously rigid and wildly uneven. The median price spread across those 375 models is exactly 0.0%: for the typical model, every provider charges the same input rate. At the same time, a substantial minority of models carry gaps large enough to change a project's cost base by an order of magnitude.

How the spreads break down

According to dev.to, the distribution is heavily concentrated at the bottom:

Input price spread	Models	Share
Under 20%	270	72%
20–50%	25	7%
50–100%	25	7%
100–200%	13	3%
Over 200%	42	11%

Nearly three-quarters of comparable models are therefore priced within a fifth of each other regardless of provider, but roughly one in nine costs more than three times as much at the most expensive provider as at the cheapest.

The largest gaps

The widest spread in the dataset belongs to Qwen/Qwen3-235B-A22B-Instruct-2507, listed at $0.09 per million input tokens on deepinfra and $3 on crusoe — a 33x difference for the same model. Other notable cases from the dev.to table:

Model	Providers	Cheapest	Priciest	Spread
Qwen3-235B-A22B-Instruct-2507	3	$0.09 (deepinfra)	$3 (crusoe)	+3233%
gpt-oss-120b	13	$0.03 (wandb)	$0.80 (crusoe)	+2567%
DeepSeek-R1-Distill-Qwen-14B	2	$0.07 (nscale)	$1.60 (together_ai)	+2186%
mistral-large-latest	3	$0.50 (mistral)	$8 (azure)	+1500%
Qwen2.5-Coder-32B-Instruct	3	$0.06 (nscale)	$0.80 (together_ai)	+1233%
DeepSeek-R1-0528	5	$0.25 (hyperbolic)	$3 (crusoe)	+1100%
Llama-3.3-70B-Instruct-Turbo	2	$0.10 (deepinfra)	$1.04 (together_ai)	+940%

Why the median is zero

dev.to attributes the uniform pricing to market structure rather than coordination. Most inference providers, the article argues, resell compute sourced from the same upstream GPU suppliers, run the same model weights on the same hardware generation, and often mirror the list price set by the original vendor or a dominant marketplace. When the underlying cost is identical and the delivered product is indistinguishable, margins compress and providers converge on a single number.

Where the outliers come from

The large gaps come from two business models at opposite ends of the spectrum. The first is GPU-hour hosts such as crusoe and hyperbolic, which rent raw compute rather than sell inference; their effective per-token price depends on utilization, power costs and hardware generation, so it can land far from commodity rates. The second is specialised platforms — azure, together_ai, nscale — that operate their own serving stacks and price in enterprise guarantees, support and integration. Mistral charging $0.50 for mistral-large-latest on its own platform while Azure charges $8 reflects two different service wrappers around the same weights.

Caveats on the data

The dev.to analysis flags three limits. Prices are point-in-time snapshots in a market where list rates change frequently, so today's outlier may disappear next week; the zero median is robust to this, the tail is not. The quoted figures are input rates at short context, and long-context requests can cost far more at some providers. And a shared model name does not guarantee identical serving: providers may use different quantization levels or batch sizes, which can legitimately justify a lower price. The article also cites 382 models in one passage and 375 elsewhere, though the analysis otherwise relies on the 375-model sample.

Why it matters

For engineering teams, the practical lesson is that one provider's price list is not the market price. For popular open-weight models, comparing endpoints can cut input costs dramatically — but only after confirming the cheaper option serves the same quantization, prices long contexts comparably, and meets reliability requirements. For the industry, the zero median suggests that much of what looks like competition in LLM inference is resale of the same upstream capacity, while the extreme spreads mark genuinely different products sold under a single model name.