You can see it on any pricing page. DeepSeek charges $0.435 per million input tokens and $0.87 per million output tokens; GPT-5.5 charges about 12 times that. When your team routes real production traffic, that gap is the difference between a bill your business absorbs and one it has to justify. This article tests whether cheaper actually means worse.
Open-weight models (DeepSeek, Qwen, Kimi) ship downloadable weights you can self-host, fine-tune and use commercially, subject to the licence and IP implications of each model; closed models (GPT-5, Claude) expose only an API. Chinese labs cut the cost of intelligence by 90 to 97 percent through the 2025 to 2026 price war, and open Chinese models now run 60 to 90 percent cheaper per token than the leading US models. The first tracks the price of intelligence; the second the per-token gap. It is the most visible edge of a broader open-versus-closed fracture. By the end you’ll know why, how it closed, and what the data proves.
Why are Chinese AI models so much cheaper than US closed models?
Two things make Chinese models cheaper, and both are structural: Mixture-of-Experts (MoE) architecture and open-weight distribution. MoE means DeepSeek V3.2 holds 671 billion parameters but activates only 37 billion per token, so each inference costs a fraction of a dense model. Open weights mean many providers host the same model and compete on price, while a closed API has one owner. DeepSeek V4 lists at $0.435 and $0.87 per million tokens.
Both trace back to one constraint: export controls denied Chinese labs the most advanced Nvidia chips, so they optimised for efficiency, and the constraint became a catalyst. Together these forces produced the 90 to 97 percent cost drop. That efficiency innovation is what broke the compute moat, the assumption that frontier AI requires ever-more Nvidia compute.
What is Mixture-of-Experts (MoE) architecture and why does it matter for model cost?
That gap starts with the architecture: a MoE model is split into specialised expert subnetworks, and only a small set activates for any token. Total knowledge decouples from per-token compute, so you get a 671-billion-parameter model’s capacity while paying for a slice of it.
FP8 (8-bit floating-point) mixed-precision training roughly halved memory bandwidth, and DeepSeek Sparse Attention cut long-context compute by about half. Sparse activation makes capacity cheap to scale.
Why did DeepSeek R1’s release wipe roughly $589 billion off NVIDIA’s market cap?
The market first priced this efficiency shift in January 2025, when DeepSeek R1 landed trained for about $5.5 million on 2,000 Nvidia H800 GPUs versus 16,000 H100s for a comparable US model. About $589 billion came off Nvidia’s market capitalisation in a day.
The sell-off repriced the compute moat. Efficiency replaced raw chip scale as the lever that moves the frontier, and demand continued to grow. Nvidia later recovered to a $5.14 trillion valuation as cheaper tokens expanded consumption, the Jevons effect, where cheaper output increases total demand.
How did Chinese open-weight labs overtake US proprietary models in twelve months?
That repricing was the opening move in a twelve-month sequence. DeepSeek V3 arrived late in 2024, then R1. GPT-5 and Llama 4 underperformed. Llama 4 Scout matched GPT-4o mini on one benchmark, a 109-billion-parameter model scoring the same as a 1.5-billion model, and landed with a thud. GPT-5 launched in August 2025 with DeepSeek V3.2 already outscoring it on AIME 2025 (96.0 to 94.6), and Wired ran “So Long, GPT-5. Hello, Qwen”.
Into that gap came Qwen, now past a billion downloads, and Moonshot’s Kimi, which opened K2 at a $4.6 million training cost. Knowledge distillation helped close the capability gap cheaply. By May 2026, Chinese open-weight models carried a majority of workloads.
What does OpenRouter token-share data show about who is winning model adoption?
That majority claim is testable against neutral routed usage. Token-share measures actual usage, not downloads. OpenRouter, the largest neutral router, moves more than 20 trillion tokens a week, and Chinese open-weight families climbed from about 2 percent to 61 percent of tokens in eighteen months.
Xiaomi’s MiMo leads vendors at about 21 percent of tokens and 22 percent of coding traffic. DeepSeek sits near 17.6 percent, ahead of every US lab, while Google’s Gemini has fallen from about 37 percent to 13 percent and Llama is below 1 percent. Check OpenRouter’s State of AI and the Artificial Analysis leaderboards.
How do Chinese open-weight and US closed models compare on cost and capability?
Adoption answers who; cost and capability answer why. A NIST/CAISI evaluation, summarised by CSIS, put DeepSeek V4 Pro about eight months behind the US frontier, not years. GLM-5.2 sits first among open-weight models on Artificial Analysis’ Intelligence Index, and Kimi K2.6 ties GPT-5.5 on SWE-Bench Pro at about 80 percent lower cost.
Against a 60 to 90 percent cost advantage (DeepSeek V4 at $0.435 and $0.87 per million tokens versus GPT-5.5 about 12 times higher), the value equation flips. Some enterprises have moved: Lindy shifted all its traffic from Claude to DeepSeek and saved millions. US closed models still hold the quality ceiling and product lock-in. The comparison splits into Qwen against Llama and DeepSeek V3 against GPT-5.
DeepSeek V3 vs GPT-5: which is better for my use case?
DeepSeek V3 wins on price-performance and control; GPT-5 holds the quality ceiling. On AIME 2025, DeepSeek V3.2 outscored GPT-5 High 96.0 to 94.6 percent, while trailing on LiveCodeBench novel code generation at 83.3 to 84.5 percent. It lists at $0.435 and $0.87 per million tokens, and its MIT-licensed open weights let you self-host and fine-tune, where GPT-5 is API-only.
High-volume, cost-sensitive and data-sovereign workloads fit DeepSeek. GPT-5 earns its premium on the hardest reasoning and its product integration.
Qwen vs Llama: which open model family should I build on?
The open-model leader has inverted. Qwen has passed a billion downloads and anchors more than 200,000 tagged models on Hugging Face, while Llama has fallen below 1 percent of OpenRouter volume. Qwen also feeds Alibaba Cloud’s AI business, an enterprise support story Llama can’t match.
For most teams, Qwen is the stronger default on price-performance and momentum. Llama still matters where US-ecosystem or compliance constraints apply. If licence terms decide it for you, check the specific open-weight licences.
The cost gap came from sparse MoE inference and open-weight distribution attacking both sides of the cost equation: less compute per token, plus multi-provider price competition. That broke the compute moat that justified closed-API pricing. The $589 billion Nvidia wipe and the 2 to 61 percent token-share shift are the same story told twice.
What the US still holds is the quality ceiling and product lock-in. The sensible default is to route to the cheapest good-enough model and reserve the frontier for the hardest tasks, instead of assuming a higher price buys more capability. Those are the geopolitical stakes behind these price moves to watch, and the cost data should now become a model-selection decision.
Frequently Asked Questions
Is it true that Chinese models are cheaper because they are worse?
No. The price gap is a structural efficiency advantage, not a discount for inferior models. Sparse MoE inference activates only a fraction of parameters per token, cutting compute, while open weights force multiple providers to compete on price. On the benchmarks that matter for most work, Chinese open-weight models now sit within months or a few percentage points of the US frontier, not years behind it.
Are “open-weight” models the same as open-source software?
Not exactly. “Open-weight” means the trained model weights are downloadable under a permissive licence, so you can self-host, fine-tune and use them commercially. It does not usually mean the training data, code and methodology are all public, which is what strict open-source licences would require. The practical difference for builders is control and cost, not full transparency.
Can I use DeepSeek or Qwen commercially, or do their licences restrict business use?
Yes, for most of the major releases. DeepSeek and Qwen ship under licences that permit commercial use, self-hosting and fine-tuning, which is why enterprises can run them privately or build products on them. The exact terms vary by model and version, so check the specific licence before a production deployment, especially where data sovereignty or redistribution obligations are a concern.
Do I need my own GPUs to run these models, or is an API enough?
You do not need your own hardware. You can call hosted APIs from providers that serve these models at open-weight prices, or use a neutral router like OpenRouter to switch between them. The advantage of open weights is the option to self-host later on your own or rented GPUs when data sovereignty, fine-tuning or predictable serving costs matter more than convenience.
Do Chinese models perform worse on English tasks because they were trained mostly on Chinese data?
No. These models are trained on large multilingual and English datasets, and they now match or beat US closed models on English coding and reasoning benchmarks such as SWE-Bench and LiveCodeBench. Kimi K2.6 ties GPT-5.5 on SWE-Bench Pro, for example. Any early language gap has largely closed, which is why English-language developers route so much production traffic to them.
Did Chinese labs just copy or distill US models to get these results?
Not simply. Distillation, where a smaller model learns from a larger teacher, is one tool in the stack, but it does not explain the price gap. The decisive work is original engineering: sparse MoE design, DeepSeek Sparse Attention and FP8 mixed-precision training. Those efficiency gains are what cut compute cost, and they hold regardless of where a model’s training signal came from.
What does the per-million-token pricing actually mean in real money for a typical app?
A million tokens is roughly the length of several long books or a large codebase’s worth of text. At DeepSeek V4’s listed $0.435 per million input and $0.87 per million output tokens, an app generating 100 million output tokens a month would cost about $87, while GPT-5.5 pricing roughly 12 times higher would push the same volume past $1,000. The gap compounds quickly at production scale.
Should I stop using GPT-5 and Claude entirely now?
No. The rational default is workload routing, not a wholesale switch. Send high-volume, cost-sensitive and self-hosted tasks to Chinese open-weight models, where the price-performance gap is decisive. Reserve GPT-5 and Claude for the hardest reasoning, safety-critical work and integrated-product features where the US frontier still holds the ceiling and the lock-in is already built into your stack.
Is Nvidia’s business at risk because efficient models need fewer chips?
Not straightforwardly. The January 2025 sell-off was the market repricing the idea that frontier AI needs ever more Nvidia compute, not a demand collapse. Since then cheaper tokens have expanded total consumption, the Jevons effect, and Nvidia recovered to a roughly $5.14 trillion valuation. The real shift is that buying more chips no longer guarantees frontier leadership, not that chips stop mattering.
What happens to these cost savings if export controls tighten again?
They would be pressured but not erased. Export controls pushed Chinese labs toward efficiency in the first place, and that engineering advantage now compounds through MoE, sparse attention and FP8 training. Tighter chip restrictions could slow flagship scaling, but the cost gap is rooted in architecture and open-weight competition, not just hardware access, so it would likely persist even under renewed constraints.
Which Chinese open-weight model should I start with for a coding workload?
Start with DeepSeek or a Qwen coding variant for general use, since both combine strong coding benchmarks with low per-token prices and easy API or self-hosted access. If you need the highest agentic coding scores, test Kimi’s latest release, which ties GPT-5.5 on SWE-Bench Pro. Route real tasks through OpenRouter and let actual coding traffic, not the leaderboard, decide.
How do I keep my data private if I self-host a Chinese model?
Self-hosting keeps your prompts and outputs inside your own infrastructure, which removes the API provider from the data path. Run the model in a private VPC, enforce strict access controls and audit the serving stack, and check the model’s licence for any telemetry or logging requirements. For regulated data, treat the model like any other third-party dependency and apply your standard security review.
