In the span of one week in late May 2026, two things happened. DeepSeek made its 75% V4-Pro discount permanent. Xiaomi slashed MiMo V2.5 Pro prices by up to 99%. Meanwhile, OpenAI launched GPT-5.5 at double the output price of its predecessor, and Anthropic shipped Claude Opus 4.7 with a new tokenizer that inflates actual costs by up to 35%.
The result: Chinese frontier AI models now cost 15-34x less than their American counterparts for equivalent capability. For cached workloads, the gap is over 100x.
This is not a temporary promotion. These are permanent prices backed by architectural innovations that make them sustainable at break-even margins.
The pricing gap in one table
| Model | Lab | Input/M | Output/M | Cache Hit/M | SWE-bench Verified |
|---|---|---|---|---|---|
| DeepSeek V4-Pro | DeepSeek | $0.435 | $0.87 | $0.003625 | 80.6% |
| MiMo V2.5 Pro | Xiaomi | $0.435 | $0.87 | $0.0036 | 79.2% |
| MiniMax M2.7 | MiniMax | $0.30 | $1.20 | N/A | ~79% |
| Kimi K2.5 | Moonshot AI | $0.60 | $2.50 | N/A | 76.8% |
| GLM-5.1 | Z.AI | $0.50 | $2.00 | N/A | 78.4% |
| Claude Opus 4.7 | Anthropic | $5.00 | $25.00 | $0.50 | 80.8% |
| GPT-5.5 | OpenAI | $5.00 | $30.00 | N/A | 78.1% |
| Gemini 2.5 Pro | $1.25 | $10.00 | $0.315 | 76.3% |
Look at the SWE-bench column. DeepSeek V4-Pro scores 80.6% — within 0.2 points of Claude Opus 4.7’s 80.8%. The output price difference: 34x. These are not budget models trading quality for cost. They are frontier models that happen to be cheap.
How we got here
The Chinese AI pricing war escalated in three phases:
Phase 1 (January-March 2026): DeepSeek V3 launched at $0.27 input / $1.10 output. Already 10x cheaper than GPT-4o. The market noticed but dismissed it as a loss-leader.
Phase 2 (April 2026): Four Chinese frontier models shipped in a 12-day window. DeepSeek V4, MiMo V2.5 Pro, Kimi K2.5, and GLM-5.1 all launched under one-third of Opus 4.6’s per-token cost. The “loss-leader” narrative collapsed — too many labs, too consistently cheap.
Phase 3 (May 2026): DeepSeek made its 75% discount permanent on May 22. Xiaomi followed on May 26 with a 99% cache hit reduction. These are not introductory offers. The labs publicly stated they break even at these prices.
Why Chinese models are structurally cheaper
Three factors explain the gap, and none of them are “they’re subsidized by the government”:
1. Architectural efficiency
DeepSeek V4 uses interleaved attention — one type compresses every 4 tokens for selective attention, another collapses every 128 tokens for global context. At 1M tokens of context, V4-Pro’s KV cache is 10% the size of a standard transformer. Single-token inference runs at 27% of the compute cost.
MiMo V2.5 Pro uses a 1:7 sparsity ratio between global and sliding window attention layers. A 70-layer model behaves like a 10-layer model for cache purposes. Storage and processing costs drop 80%.
These are not tricks. They are genuine architectural innovations that reduce the compute needed per token.
2. Hardware cost differences
Chinese labs operate primarily on domestic hardware (Huawei Ascend 910B, custom inference chips) with lower per-unit costs than NVIDIA H100/H200 clusters. Energy costs in Chinese data center regions run 30-50% below US equivalents.
3. Competition dynamics
Five major Chinese AI labs (DeepSeek, Xiaomi, Alibaba/Qwen, Moonshot/Kimi, Z.AI/GLM) are competing for the same developer market. Price is the primary differentiator when benchmark scores converge. American labs face less direct price competition — OpenAI, Anthropic, and Google compete more on features and ecosystem lock-in.
Meanwhile, American labs raised prices
The contrast is stark:
- GPT-5.5 (April 2026): Output price doubled from $15/M to $30/M compared to GPT-5.4
- Claude Opus 4.7 (May 2026): Same rate card ($5/$25) but a new tokenizer produces up to 35% more tokens for the same text, effectively raising costs
- Gemini 2.5 Pro: Held steady at $1.25/$10 — the cheapest American option, but still 11x more expensive than DeepSeek on output
The American labs are betting that developers will pay a premium for ecosystem integration (ChatGPT plugins, Claude’s computer use, Gemini’s Google Cloud integration). The Chinese labs are betting that price wins when quality is equivalent.
Real-world cost comparison
What does a typical developer workload actually cost?
| Workload | GPT-5.5 Cost | Claude Opus 4.7 | DeepSeek V4-Pro | MiMo V2.5 Pro |
|---|---|---|---|---|
| 1hr coding agent session | $8-15 | $10-20 | $0.30-0.60 | $0.25-0.50 |
| Process 1000 documents (RAG) | $150-300 | $125-250 | $4-8 | $4-8 |
| Run 100 SWE-bench tasks | $500+ | $400+ | $15-25 | $12-20 |
| Monthly agent pipeline (24/7) | $5,000-10,000 | $6,000-12,000 | $200-400 | $150-350 |
For a startup running an AI-powered product, the difference between $10,000/month and $300/month is the difference between burning runway and being profitable.
The catch (there is always a catch)
Chinese models are not a drop-in replacement for every use case:
- Data residency: API calls route through Chinese infrastructure. For regulated industries (healthcare, finance, government), this may be a compliance issue.
- Latency: Depending on your location, latency to Chinese API endpoints may be 50-200ms higher than US-based alternatives. For real-time chat, this matters. For batch processing, it does not.
- Ecosystem: No native integrations with AWS, Azure, or GCP. You use the OpenAI-compatible API directly or go through OpenRouter.
- Support: Documentation is improving but still thinner than OpenAI or Anthropic. Community support is growing fast.
- Availability: Some Chinese APIs have experienced brief outages during peak demand. Uptime is generally 99.5%+ but not yet at the 99.99% level of major US providers.
For most developer workloads — coding agents, document processing, content generation, RAG pipelines — none of these are dealbreakers. For a full list of the best Chinese models ranked by capability, see our best Chinese AI models guide.
How to get started
The fastest path to using Chinese models:
- OpenRouter — Single API key, access to DeepSeek V4-Pro, MiMo V2.5 Pro, Kimi K2.5, and more. Small markup but unified billing. Setup guide.
- Direct API — Create an account at the provider’s platform. All use OpenAI-compatible endpoints. MiMo API guide | DeepSeek API guide.
- Coding tools — Both models work with Aider, Claude Code, Continue, and any tool that supports custom OpenAI-compatible endpoints.
- Full migration — If you are currently on GPT-5.5 or Claude, follow our step-by-step migration guide with eval framework, router pattern, and fallback handling.
FAQ
Are Chinese AI models safe to use for commercial projects?
Yes. DeepSeek, MiMo, and Kimi all have commercial licenses. The models are open-source (or open-weight), and the APIs have standard terms of service. Check each provider’s specific license for your use case.
Will American labs match these prices?
Unlikely in the short term. Their cost structures (NVIDIA hardware, US energy prices, higher headcount) make sub-$1/M output pricing difficult without architectural changes similar to what Chinese labs have implemented.
Which Chinese model should I pick?
For coding: DeepSeek V4-Pro (highest SWE-bench) or MiMo V2.5 Pro (best token efficiency). For general tasks: MiniMax M2.7 or Kimi K2.5. For reasoning: DeepSeek V4-Pro. See our MiMo vs DeepSeek comparison for a detailed breakdown.
Is the 99% cache discount real or marketing?
Real. Xiaomi’s Fuli Luo published the technical explanation: hierarchical KV cache optimization for SWA reduces cache storage by 80%. The inference engine breaks even at these prices. See our MiMo price cut breakdown for the full technical details.
How do benchmarks compare at these price points?
DeepSeek V4-Pro (80.6% SWE-bench) matches Claude Opus 4.7 (80.8%) within margin of error. MiMo V2.5 Pro (79.2%) beats GPT-5.5 (78.1%). You are not sacrificing quality for price — you are getting equivalent quality at 15-34x lower cost.