US AI models are often cheaper per task than Chinese rivals, analysis finds

According to a Seeking Alpha report published October 6, 2026, which draws on data from Artificial Analysis, US models from OpenAI, Anthropic and Google often cost less per completed task than Chinese competitors. A cheaper sticker price, it turns out, doesn’t guarantee a cheaper result.

Price per token versus price per job

In AI, the unit of billing is a token — a small chunk of text a model reads or writes.

Chinese models from DeepSeek, Alibaba’s Qwen and Moonshot’s Kimi do charge less per token. Claims about that advantage put Chinese per-token pricing at 60-90% lower than US systems, though the report contests that framing.

The analysis finds US models tend to be more efficient on that measure. They reach the answer using fewer tokens, which can bring the total bill for a completed task below what a Chinese model would cost.

The report features insights from Martin Chorzempa of the Peterson Institute. Its overall argument is that the gap between US and Chinese AI costs is narrower, and sometimes reversed, compared with what headline pricing suggests.

The hidden toll on “free” models

Many Chinese models are released as open-weight systems, meaning anyone can download them for free.

Per the analysis, commercial use through cloud providers can involve cuts of up to 30% going to those providers. That raises the effective cost for enterprises relying on cloud infrastructure rather than running the models locally.

Enterprise AI spending is cooling, even as usage climbs

Data from Ara Kharazian, lead economist at Ramp, shows enterprise AI spending fell 5.2% in recent tracking. That decline came even as token consumption hit record highs.

The research frames this as a stabilization of spending patterns. It counters fears that companies are simply throwing ever-larger checks at AI vendors with no ceiling in sight.

Why the narrative matters

This analysis doesn’t claim Chinese models are expensive. It argues that the comparison most people make, price per token, captures only part of the picture.

Measuring cost per completed task tells a more complicated story. On that metric, US models often come out ahead.

What this means for buyers and investors

For enterprise buyers, the practical takeaway is to benchmark by the job, not the token. Cloud hosting fees belong in that tally too, especially for open-weight models accessed through third-party platforms where cuts of up to 30% can apply.

The Ramp data adds another layer. A 5.2% decline in spending alongside record usage points to an AI market that may be entering a more mature phase, where cost-effectiveness matters as much as adoption.