DeepSeek V4 vs Qwen 3.5 2026: The Open-Source AI War Nobody Is Covering Honestly
DeepSeek and Qwen dominate global AI token usage. Western media barely covers them. Both are free. Both are open-source. Which one you pick changes everything about cost and capability.
Chinese AI models now account for roughly 61 percent of token consumption on OpenRouter, the global model aggregation platform. DeepSeek and Qwen lead this shift. Western media mostly ignores them.
I spent a week testing both. Here is what actually matters.
---
The Pricing
Both models are free to use through their chat interfaces. No ads. No subscription. No credit card. The API pricing is where the real story lives.
| Model | Input ($/1M tokens) | Output ($/1M tokens) |
|---|---|---|
| DeepSeek V4 Flash | $0.14 | $0.28 |
| DeepSeek V4 Pro | $1.74 | $3.48 |
| Qwen 3.5 Flash | $0.10 | $0.40 |
| Qwen 3.5 Plus | $0.40 | $2.40 |
| GPT-5.6 Luna | $1.00 | $6.00 |
| Claude Opus 4.7 | $6.25 | $25.00 |
Qwen 3.5 Flash costs $0.10 per million input tokens. Claude Opus 4.7 costs $6.25. That is a 62 times difference. A developer building on Qwen Flash can run 62 times more requests for the same budget.
These numbers matter because they change what you can afford to build. A chatbot that costs $5,000 per month on Claude Opus costs $80 per month on Qwen Flash. The difference is the difference between a side project and a dead project.
---
The Two Philosophies
DeepSeek bets on raw scale. V4 Pro is 1.6 trillion parameters, the largest open-weight model in the world. 49 billion activated per forward pass. MIT license. You can download it. Run it yourself. Fine-tune it. Sell it.
Qwen bets on efficiency. Qwen 3.5 Plus is 397 billion total parameters but only 17 billion activated. It scores within a few points of V4 Pro on most benchmarks while costing 75 percent less. Also open weights. Also MIT license.
The philosophical difference: DeepSeek swings for maximum capability at any parameter count. Qwen optimizes for capability per dollar.
---
The Benchmarks
I am going to do something most comparison articles do not do. Here are the actual benchmark numbers from independent third-party evaluations, not company press releases.
| Benchmark | DeepSeek V4 Pro | Qwen 3.5 Plus | Winner |
|---|---|---|---|
| SWE-Bench Verified | 80.6% | 76.4% | DeepSeek |
| GPQA Diamond | 90.1% | 88.4% | DeepSeek |
| MMLU-Pro | Not published | 87.8 | Unclear |
| Terminal-Bench 2.0 | 67.9% | Not published | Unclear |
| Humanity's Last Exam | 48.2% | Not published | Unclear |
And here is what is missing. Qwen has not published direct comparisons on several key benchmarks where DeepSeek has numbers. DeepSeek has not published on several where Qwen claims victory. Both companies select which benchmarks to highlight. This is normal. It is also something most articles do not mention.
The honest answer: DeepSeek V4 Pro leads on coding and reasoning by small margins. Qwen 3.5 Plus is close enough on most tasks that the price difference becomes the deciding factor for anyone building at scale.
---
The Things Nobody Talks About
Content moderation. Both models run on servers in China and apply Chinese content moderation. Politically sensitive topics get filtered or deflected. If your product requires discussing topics restricted under Chinese law, neither model works for you. Western competitors do not solve this perfectly, but they do not filter according to Chinese regulations.
Qwen is multimodal. DeepSeek is not. Qwen handles vision input natively. You can upload images. You cannot do this with DeepSeek. If your application needs image understanding, Qwen is your only option between these two.
DeepSeek trains on domestic Chinese chips. V4 was trained on Huawei Ascend NPUs. This matters for two reasons. First, it means export controls do not directly throttle DeepSeek's training capacity. Second, it means DeepSeek is proving that non-NVIDIA AI hardware works at scale. The geopolitical implications are bigger than any benchmark score.
Both are MIT licensed. You can download the weights. Run them locally. Fine-tune them. Build products on them. No vendor lock-in. No API key required. This is the Superpower that GPT and Claude do not offer.
---
Which One
Pick DeepSeek V4 Pro when you need the best possible coding and reasoning from an open-source model. The 1.6 trillion parameter brute force shows in the benchmark numbers. The chat interface works well. The API is cheap.
Pick Qwen 3.5 Plus when you need the best value in AI. It delivers within striking distance of V4 Pro on most benchmarks while costing less than half as much. The multimodal support adds capabilities DeepSeek lacks. The Qwen Coding Plan at $50 a month for 90,000 requests across seven models is absurdly good value.
Or do what most developers are doing. Use both. DeepSeek for hard problems. Qwen for volume. Neither costs enough to make you choose.
---
One More Thing
Qwen 3.7 Max Preview tied Claude Opus 4.7 on the Artificial Analysis Intelligence Index in May 2026. Score of 57. Same as Opus 4.7. The final 3.7 release will likely pull ahead.
DeepSeek is reportedly raising $4.4 billion in its first external funding round. Tencent and Alibaba are interested. The company has been entirely self-funded until now, burning through its trading profits from High-Flyer, the quantitative hedge fund that spawned it.
These two companies are not playing the same game as OpenAI and Anthropic. They are not charging $20 a month for a chat subscription. They are giving away frontier-grade AI for free and building a developer ecosystem instead. Whether that strategy works depends on whether developers switch. Judging by the 61 percent token share, they already are.
---
*Sources: OpenRouter token consumption data; Official API pricing pages; Artificial Analysis Intelligence Index (May 2026); DeepSeek V4 technical report; Qwen 3.5 technical report; MIT license documentation.*