Why Chinese AI Models Cost Less Than OpenAI or Anthropic

The price gap between Chinese AI models and Western ones like OpenAI and Anthropic is pretty significant — and the instinct for any budget-conscious marketer is to ask why they're paying more. Let's reason through it.

Key Points

• U.S. companies' use of Chinese AI models via OpenRouter has stayed above 30% of weekly tokens since early February, peaking at 46%, up from an 11% average the year before, according to CNBC.
• Z.ai's GLM 5.2 saw daily token volume grow roughly 27x and its customer count grow about 80x in its first week after launch, per Vercel.
• Chinese open-weight models can run 60% to 90% cheaper than leading Anthropic and OpenAI systems while landing within a percentage point of top benchmarks on some agentic tasks.
• AI startup Lindy moved 100% of its traffic from Claude to DeepSeek and expects to save millions of dollars within months.
• The real risk isn't which country builds the better model. It's that companies who built their entire AI stack around one vendor now face a costly, disruptive migration every time pricing or capability shifts.

What's Driving Lower Prices From Models Like DeepSeek

We want to be careful here, because this story gets told two ways, and only one of them is actually useful. The exciting version says Chinese AI models are catching up to American frontier labs. The useful version says a lot of U.S. companies built their AI strategy on the assumption that one vendor's pricing would stay reasonable forever, and that assumption just broke.

The Numbers Behind the Shift

CNBC reported that U.S. companies' share of tokens routed to Chinese models through OpenRouter has held above 30% every week since February 8, touching 46% at points. A year earlier, that average sat at 11%, and as recently as the first half of 2025 it was 4.5%. That's not a gradual drift. That's a market repricing its assumptions in real time.

The trigger is straightforward: token prices at leading U.S. labs have climbed as capability has climbed, and companies that once adopted AI regardless of cost are now, as Brookings fellow Kyle Chan told CNBC, getting "more cost-conscious." Meanwhile, models like DeepSeek's V4 and Z.ai's GLM 5.2 are closing the performance gap while staying dramatically cheaper. GLM 5.2 landed within a percentage point of Anthropic's Opus 4.8 on a closely watched agentic benchmark, at roughly a fifth of the cost, and grew its customer base 80x in its first week on Vercel.

Why This Is a Cost Story, Not Just a Capability Story

Lindy's move is the clearest example. The company shifted 100% of its traffic from Claude to DeepSeek and watched, in CEO Flo Crivello's words, its cost curve "crash to the ground." OpenRouter's Justin Summerville puts the gap at 60% to 90% cheaper for open-weight Chinese models doing comparable work. When a model that's six to nine months behind frontier capability, by Chan's estimate, costs a fraction of the price, a lot of workloads don't need the frontier model at all. They need the cheapest model that clears the bar.

That's a rational decision at the individual company level. It's also exactly the kind of decision that creates concentration risk at scale, just pointed in a different direction than people expected two years ago.

The Single-Vendor Trap Nobody Priced In

Here's the part we think gets undersold. Hugging Face's Yacine Jernite told CNBC there's a real risk that companies end up choosing between expensive U.S. proprietary models with volatile pricing, or Chinese open-weight models as the only viable way to control costs. Both of those are single-dependency positions. Neither is actually a strategy.

Companies that built products entirely on one closed-source American model are now discovering their unit economics were never really theirs to control. The lesson isn't "switch to the cheaper option." It's that any AI stack with one point of dependency, regardless of which country built the model, carries a financial risk that doesn't show up until pricing moves. Building an AI-informed growth strategy that assumes today's token price holds for the next three years is a bet, not a plan.

What This Means for Marketing and Growth Teams

If your team is running campaigns, content generation, or customer-facing agents on a single model provider, this is worth an honest cost audit now, not after the next price increase. Some workloads genuinely need frontier reasoning. Most don't. Knowing which is which, and building the flexibility to route accordingly, is the actual skill here, not picking a geopolitical side. We help clients build AI marketing infrastructure that's resilient to exactly this kind of vendor shock, because it's going to keep happening.

CNBC's reporting on this shift is worth reading in full if you're making a model decision for your team this quarter.

If your AI stack has one point of failure and you'd rather find out about it now than during your next invoice, that's a conversation worth having with us.