Same Model, 3x the Price: Why LLM Token Pricing Varies So Much

Run the same model, with the same weights, through two different API providers and you can pay up to 3x more per million tokens at one than the other. DeepSeek V3.2, one of the most widely hosted open-weight models in production, has traded between roughly $0.18 and $0.57 per million input tokens across providers this year. Identical model, identical output, three times the price.

GPU buyers will recognize this pattern. We measured the same thing on hardware: a 2.5x spread on H100 rentals and a 7.7x spread on B200. Token pricing inherits the fragmentation of the compute market underneath it, then adds a few distortions of its own.

This piece breaks down where the token spread comes from, what a fair price actually is, and how buyers should compare providers when the sticker price is only part of the story.

The spread, measured

Public rate cards for DeepSeek V3.2 across major hosting providers:

Two observations before the why. First, the official model developer is not the cheapest host; several third parties undercut DeepSeek's own API. Second, even this table understates the real spread, because list price is not effective price: cache discounts, quantization, and context limits change what you actually pay per unit of work.

The pattern is not unique to one model. Moonshot's Kimi K2.6 lists at $0.95 input and $4.00 output on the official API while third-party hosts quote it at $1.20 and $4.50. Every widely hosted open-weight model shows a version of this spread.

Why identical tokens price differently

1. The hardware underneath prices differently. Serving cost is GPU economics. A provider running owned H200s at high utilization has a cost base near $2.30 per GPU-hour; a provider renting hyperscaler capacity on-demand pays several times that. The cost-per-token chain runs from GPU-hour cost through throughput to price per million tokens, so different infrastructure cost bases produce different token floors before margin even enters.

2. Utilization is the hidden variable. Inference margins live and die on batch density. A provider running fleets at 80 percent utilization can price near cost and win on volume. One running at 40 percent needs double the margin per token to cover the same hardware. Buyers never see this number, but it explains much of why aggressive newcomers undercut established hosts.

3. Not every endpoint serves the same product. Some budget endpoints run FP8 or otherwise quantized versions of the model. For many workloads the quality difference is negligible; for some it is not. A cheaper price for a quantized endpoint is not the same price for the same good, and most price comparisons ignore this entirely.

4. Cache pricing changes the effective price more than list price does. DeepSeek's official API charges $0.028 per million cached input tokens, a 90 percent discount. A workload with heavy prompt reuse can be cheaper on a higher list price with strong cache terms than on the lowest sticker price without them. Two providers with identical list prices can differ 2x in effective cost on a real workload.

5. Margin strategy and market position. Some hosts price tokens as the product. Others price them as a loss leader for a platform. Hyperscaler-hosted open-weight endpoints typically anchor the top of the range, priced for enterprises that value the procurement relationship over the rate. This is the same sales-overhead-and-margin structure we documented in GPU rental pricing, reproduced one layer up.

What is a fair token price?

Fair price for an open-weight model is answerable in a way it never was for closed models, because anyone can compute the serving cost floor. The chain: hardware cost per GPU-hour, times GPUs required, divided by realized throughput, gives cost per million tokens. Add sustainable margin and you have what a competitive market should converge to.

The spread persists because that convergence has nowhere to happen. Token prices are posted, not traded. There is no venue where supply competes on a standard product at visible prices, so each provider prices to its own segment and the gaps sit unarbitraged. This is the exact structure we described in How a Token Exchange Works: posted prices without a market mechanism produce persistent dispersion that a functioning exchange would compress.

The GPU market is a preview. H100's cross-provider spread has narrowed as the rental market matured and price visibility improved. B200's 7.7x spread is what immaturity looks like. Token pricing in 2026 sits closer to the B200 end: young, opaque, and wide.

How buyers should actually compare

• Compare blended cost on your token mix, not input list price. Weight input, output, and cached tokens the way your workload actually consumes them. The Mercatus Token Index computes a Standard Blended Price per model for exactly this reason.
• Confirm the precision. Ask whether the endpoint serves full-precision or quantized weights, and benchmark quality if your workload is sensitive.
• Price the cache terms. For agent and RAG workloads with heavy prompt reuse, cache discount structure routinely outweighs list price differences.
• Check context and feature parity. Some hosts cap context below the model's maximum or omit tool calling. A cheaper endpoint that cannot run your workload is not cheaper.
• Recheck monthly. These prices move constantly and rankings flip. What was the cheapest host in March is often not by August.

For a continuously updated view, the Token Index tracks blended per-model pricing across providers daily, weighted by real usage.

Frequently asked questions

Why does the same model cost more on one API than another?

Different hosts have different hardware costs, utilization rates, quantization choices, cache structures, and margin strategies. The model weights are identical; the businesses serving them are not.

Is the cheapest provider worse?

Not necessarily, but verify two things: whether the endpoint is quantized, and whether context and feature support match your workload. Some budget endpoints are identical service at lower margin; some are a different product.

Why doesn't the official model developer have the lowest price?

The developer's API is one host among many, with its own infrastructure costs and margin targets. Third parties with cheaper compute or leaner margins regularly undercut official APIs on open-weight models.

Will token price spreads narrow?

The GPU market suggests yes, as transparency and competition increase. Spreads compress when prices become visible and comparable in one place, which is the function of a price index and, eventually, an exchange with real price discovery.

What is a blended token price?

A single per-model cost metric weighting input, output, and cached tokens by realistic usage shares, rather than quoting input price alone. It is how the Token Index benchmarks models, and it is the number that predicts your actual bill.

Methodology

Provider pricing compiled from published rate cards and public pricing trackers, 2026. Blended pricing methodology follows the Mercatus Token Index Standard Blended Price. GPU cost baselines from Mercatus server cost analyses. Last verified: 2026-08-02.