Back to Blog
Industry

The China Discount Is Dead at the Top. Check Your Router.

Qwen4-Max and Gemini 4 Hydrogen now share a $1.50 input price, shrinking the frontier "China discount" to 13–25%. The 20x gap now sits one tier down in flash models, so route by tier, not by flag.

OPC Community

OPC Community

Nov 17, 2026 8 min read
The China Discount Is Dead at the Top. Check Your Router.

Qwen4-Max and Gemini 4 Hydrogen now share an input price. The cheap tokens still exist, one tier down, and most routing configs point at the wrong place.

Alibaba listed Qwen4-Max this week at $1.50 per million input tokens. Google listed Gemini 4 Hydrogen at $1.50 per million input tokens.

Same number, same week, to the cent.

If you build on LLMs, some line in your stack assumes this can't happen. Maybe it's a fallback rule. Maybe it's a comment in a config file, or a sentence in a planning doc that says "for the high-volume stuff we'll use a Chinese model, they're cheaper." That sentence held up for about three years. At the frontier tier it stopped being true on a Tuesday, and nobody sent you a memo.

Thirteen Percent Is Not a Discount

You'll object that output prices still differ, and they do. Qwen4-Max charges $4.50 per million output tokens. Hydrogen charges $7.50. Forty percent off is forty percent off.

Sure, if your bill were mostly output. Almost nobody's is anymore.

Take an agent workload, the kind most of us run now. You've got a long system prompt, tool definitions, retrieved documents, and a transcript that grows and gets fed back in on every turn. Ten input tokens for every output token is a conservative ratio for that shape. Do the arithmetic on eleven million tokens and Qwen4-Max comes to $19.50. Hydrogen comes to $22.50.

That's a 13% gap.

At three-to-one, which looks more like a chat product, the gap widens to 25%. So the honest range for the frontier-tier "China discount" today is 13 to 25 percent, and where you land depends on how much your app reads versus how much it writes.

Now add up what it costs you to collect that. You take on a second vendor relationship. You build a second set of evals, because the two models fail differently. You have the data-residency conversation with every enterprise customer who reads your subprocessor list. And the prompts you tuned for one model's habits regress on the other's.

Thirteen percent doesn't cover that. It doesn't even cover the engineer-week you'll burn discovering that your tool-calling prompt behaves differently.

The old discount was 5x, sometimes 10x, and at 5x you swallow all of that friction without thinking. At 13% the friction is the bigger number. The decision flipped, and the routing rule in your head didn't.

The Discount Was Never About China

I think most people have this wrong at the root.

We filed cheap Chinese models under geography. Lower costs, subsidized compute, a brutal domestic price war. All of that is real. But the discount at the top tier never told you much about a country. It told you about position.

You price below the leader when you believe a buyer needs a reason to pick you. Challengers do this in every market. The discount was a concession, an admission written in the price column: we know you'd rather use the other one, here's 80% off to try ours.

Read it that way and $1.50 is the loudest thing in Alibaba's launch. Louder than the benchmark.

The benchmark is in there too. Alibaba's announcement cites an Intelligence Index score of 43. But a score is a claim about the model, and a price is a claim about the customer. By matching Google's input price, Alibaba is saying it thinks you'll choose Qwen4-Max at the same cost, with no bribe attached.

They could be wrong about that. Test it on your own tasks before you believe any index. Still, look at what kind of statement it is. A lab that saw itself as the fallback option would never price this way.

“A discount is what you charge when you think you're second.”

And the day a vendor stops charging it is the day you should stop treating them as the budget lane. That cuts both ways. Qwen4-Max no longer earns a slot in your router for being cheap. It has to earn one for being good, which is a harder audition and a more interesting one.

The Sticker Stopped Being the Price

Something else in the Qwen launch matters more than the per-token number, and it's easy to read right past it.

Alibaba led with cost per task. $3.27 to run the benchmark suite, which they say is a third of what Qwen3.8 Max cost on the same work. Part of where that came from: Qwen4-Max uses 39% fewer output tokens to get to an answer.

Think about what happened there. The bill for the same work fell by two-thirds, and a large share of the drop has nothing to do with the rate card. The model talks less. It reasons in fewer tokens, and a token it never emits costs you exactly zero.

So per-token price has become a bad proxy for what you'll spend, and it gets worse with every release. Two models at an identical $1.50 input rate can hand you bills that differ by 2x, because one of them burns through reasoning tokens like a nervous intern narrating every step.

I don't have Hydrogen's per-task figure on the same suite. I'd bet you don't either, and that's exactly the problem. Most of us compare vendors on the one number printed on the pricing page, while the vendors have quietly started competing on a different one.

The founders I see getting this wrong aren't careless. Eighteen months ago they built a spreadsheet with a column for input price and a column for output price, and the spreadsheet still looks rigorous. It measures the price of gas in a world where the cars now differ 3x in mileage.

Log tokens per completed task, per model, on your real traffic. Multiply that by the rate card and you have your price. Nothing on anyone's website will tell you what it is.

The Twenty-X Gap Is One Floor Down

The discount didn't vanish. It moved to where the volume is.

In the same week, Tencent shipped Hy4 flash at 0.5 yuan per million input tokens and 2 yuan per million output. At roughly 7.1 yuan to the dollar, that's about seven cents in and twenty-eight cents out.

Run the same ten-to-one agent workload. Eleven million tokens through Hy4 flash costs about 98 cents. Through Qwen4-Max, $19.50.

Twenty times cheaper. And that's a Chinese model undercutting a Chinese model.

This should rearrange your mental map. The big price gap in the market no longer runs between American labs and Chinese labs. It runs between tiers. The line that matters is horizontal now, and it sits between "frontier" and "flash" on every vendor's menu.

At the top, everyone is converging on the same price because everyone is paying for the same scarce thing: the compute to serve a very large model to people who need the best answer. At the bottom, the models are small enough that serving cost collapses, and the Chinese labs are fighting each other for volume with pricing that still looks like the old days.

So the old instinct was half right. If you want the absurdly cheap tokens, China is still where a lot of them are. You just have to look one floor down from where you've been looking.

Route by Tier, Not by Flag

Here's what I'd change this week.

Delete the rule that says "Chinese model for cost." In its place, put two separate questions that your router asks in order.

First: does this call need a frontier model at all? Be ruthless, because this question is worth 20x and the next one is worth 13%. Classification, extraction, routing, summarizing a tool result, deciding whether a retrieved chunk is relevant. In most agent pipelines these calls outnumber the hard ones by a wide margin, and they run fine on a flash-tier model. Every one you send to a $1.50 model is a call you're overpaying for twentyfold.

Second, and only for the calls that survive the first question: which frontier model is best at this task, on my evals? I mean best, and I don't mean cheapest. With input prices level, quality and reliability decide it, and cost is a tiebreaker you settle with your own per-task numbers.

Then check your ratio. If your workload is output-heavy, say long-form generation or code that writes whole files, Qwen4-Max's $4.50 output rate is a real edge over $7.50 and worth testing for. If you're input-heavy, the edge is a rounding error and you should choose on other grounds entirely.

One caution, since I'm the one telling you to move volume down a tier. The cheap tier is where failures get quiet. When a frontier model is wrong, it tends to be wrong in an interesting way you'll notice. A flash model that's wrong returns a confident, well-formatted, slightly off classification ten thousand times a day. Put an eval on the flash calls before you move them, and sample the outputs after. The savings are real, and they evaporate just as reliably if nobody's watching.

I know the anxiety underneath all of this. Prices and model names change every few weeks, and any cost decision you make feels like it'll be stale by the time you ship it. It will be. That's why you encode the question into your router and leave the answer out. "Use the cheapest model that passes this eval" survives a pricing change. "Use Qwen, it's cheap" didn't survive this week.

Go find that comment in your config. The one that says something like # fallback: cheaper.

Whoever wrote it was describing a market where second place came with a coupon. Alibaba just stopped handing out the coupon, and they did it on purpose.

"Who's cheaper?" was always the wrong question for that comment to ask. The right one is "who still thinks they have to be?"

Frequently Asked Questions

Is Qwen4-Max still cheaper than Gemini 4 Hydrogen?

Only on output. Both charge $1.50 per million input tokens, while Qwen4-Max charges $4.50 per million output tokens against Hydrogen's $7.50. On a ten-to-one input-heavy agent workload that works out to a 13% gap, widening to 25% at three-to-one, which rarely covers the cost of a second vendor, a second set of evals, and re-tuned prompts.

Where are the cheap Chinese LLM tokens now?

One tier down, in flash-class models. Tencent's Hy4 flash costs 0.5 yuan per million input tokens and 2 yuan per million output, roughly seven cents in and twenty-eight cents out. The same eleven-million-token agent workload costs about 98 cents there versus $19.50 on Qwen4-Max, a gap of about 20x.

How should I set up LLM routing now that frontier prices have converged?

Route by tier, not by country. First ask whether a call needs a frontier model at all, since classification, extraction, and summarizing tool results usually run fine on a flash-tier model. For the calls that do need one, pick the best frontier model on your own evals and compare cost using tokens per completed task on your real traffic. Put an eval on the flash calls before you move them and sample outputs afterward, because failures at the cheap tier are quiet.

Share this article

Join the OPC Community

Connect with solo founders worldwide. Get daily insights, curated opportunities, and peer support.

Join free