Short version: The old version of this article was stale. DeepSeek V4 has been released, and Qwen3-Max-Thinking is no longer just a benchmark rumor. Checked May 10, 2026: DeepSeek V4 Flash is the best first test for cheap 1M-context API work. DeepSeek V4 Pro is the stronger DeepSeek option, but its discounted pricing is temporary through May 31, 2026. Qwen3-Max-Thinking is a stronger fit if you are already in Alibaba Cloud Model Studio, need Qwen's tool-use stack, or want Qwen's official reasoning benchmark profile.
- DeepSeek-V4 Preview went live on April 24, 2026 with V4 Pro and V4 Flash
- DeepSeek V4 supports 1M context, OpenAI-format API, Anthropic-format API, and open weights
- DeepSeek V4 Flash is listed at $0.14 cache-miss input and $0.28 output per 1M tokens
- DeepSeek V4 Pro has a 75% discount through May 31, 2026: $0.435 cache-miss input and $0.87 output per 1M tokens
- Alibaba Cloud lists qwen3-max-2026-01-23 as the thinking-mode Qwen3-Max snapshot
- Qwen3-Max supports 262,144 context, 81,920 max chain-of-thought tokens, and 32,768 max output tokens
- Alibaba Cloud Global pricing for qwen3-max starts at $0.359 input and $1.434 output per 1M tokens for requests up to 32K tokens
- Qwen's official blog reports Qwen3-Max-Thinking at 85.9 on LiveCodeBench v6, 75.3 on SWE Verified, and 49.8 on HLE with tools
This used to be a DeepSeek V3 vs Qwen benchmark post. That framing is now wrong. DeepSeek V4 is live, the old pre-release warning is false, and the useful comparison is now DeepSeek V4 Flash/Pro vs Qwen3-Max-Thinking.
This update uses official DeepSeek API docs, DeepSeek's official V4 release note, Qwen's official Qwen3-Max-Thinking blog, and Alibaba Cloud Model Studio docs. I am not treating old third-party benchmark roundups as source of truth.
Short answer
The model choice in plain English.
Choose DeepSeek V4 Flash if cost and 1M context matter most. Choose DeepSeek V4 Pro if you want the stronger DeepSeek model for harder reasoning, agentic coding, or world knowledge, and you are comfortable with the temporary-discount pricing. Choose Qwen3-Max-Thinking if you are building in Alibaba Cloud Model Studio, need Qwen's tool-use path, or want the Qwen reasoning model with official benchmark coverage against GPT-5.2-Thinking, Claude Opus 4.5, Gemini 3 Pro, and DeepSeek V3.2.
Quick recommendation
| Need | Pick | Why |
|---|---|---|
| Cheapest long-context API default | DeepSeek V4 Flash | Lowest official listed cost and 1M context |
| Harder reasoning or coding inside DeepSeek | DeepSeek V4 Pro | DeepSeek positions it as the flagship V4 model |
| Alibaba Cloud / Model Studio workflow | Qwen3-Max-Thinking | Native Qwen/Alibaba support, tool calling, and tiered Model Studio pricing |
| Open-weight experimentation | DeepSeek V4 | DeepSeek links open weights from the official release note |
| Qwen reasoning benchmark profile | Qwen3-Max-Thinking | Qwen reports strong LiveCodeBench, HLE-with-tools, and Arena-Hard v2 results |
Current status
What is actually live now.
Release and API status
| Item | DeepSeek V4 | Qwen3-Max-Thinking |
|---|---|---|
| Current status | DeepSeek-V4 Preview is live from April 24, 2026 | Qwen3-Max-Thinking announced by Qwen on January 25, 2026 |
| API model names | deepseek-v4-flash, deepseek-v4-pro | qwen3-max and qwen3-max-2026-01-23 in Alibaba Cloud docs |
| Context | 1M | 262,144 tokens |
| Max output | 384K listed by DeepSeek | 32,768 output tokens in thinking mode |
| Tool use | Tool calls supported | Alibaba docs list tool calling support |
| Weights | Open weights linked by DeepSeek | Alibaba/Qwen docs describe API/model availability; do not assume DeepSeek-style open weights |
Source: DeepSeek API Docs, Qwen blog, Alibaba Cloud Model Studio docs
DeepSeek, Qwen, Claude, or GPT? Find the model that fits your use case in 60 seconds.
12 models · Personalized picks · 60 seconds
The old article was wrong after April 24
Any pre-release text about DeepSeek V4 is now outdated. DeepSeek's own release note says V4 Preview is live, open-sourced, and available through the API.
Pricing
DeepSeek is cheaper, but watch the Pro discount date.
DeepSeek and Alibaba price differently. DeepSeek lists one price table for V4 Flash and V4 Pro, with separate cache-hit and cache-miss input pricing. Alibaba Cloud lists Qwen3-Max pricing by deployment mode and request size. The table below uses DeepSeek's official pricing and Alibaba Cloud's Global deployment pricing because that is the most relevant public non-China deployment mode in the docs.
Official API pricing snapshot, checked May 10, 2026
| Model | Input price | Output price | Notes |
|---|---|---|---|
| DeepSeek V4 Flash | $0.0028 cache hit / $0.14 cache miss per 1M tokens | $0.28 per 1M tokens | 1M context |
| DeepSeek V4 Pro | $0.003625 cache hit / $0.435 cache miss per 1M tokens | $0.87 per 1M tokens | 75% discount through May 31, 2026; list output is $3.48 |
| Qwen3-Max Global, <=32K input | $0.359 per 1M tokens | $1.434 per 1M tokens | Alibaba Cloud tiered Global pricing |
| Qwen3-Max Global, 32K-128K input | $0.574 per 1M tokens | $2.294 per 1M tokens | Higher tier for longer requests |
| Qwen3-Max Global, 128K-252K input | $1.004 per 1M tokens | $4.014 per 1M tokens | Highest listed Global tier in the source |
Source: DeepSeek Models & Pricing; Alibaba Cloud Model Studio pricing
The practical reading: DeepSeek V4 Flash is the cheaper default. Qwen3-Max is not priced like a bargain basement model once you use longer inputs. It can still be the right choice if you need Qwen's ecosystem, languages, tool path, or benchmark profile.
Benchmarks
Provider claims, not independent proof.
Qwen publishes more text benchmark detail for Qwen3-Max-Thinking than DeepSeek's text release page exposes for V4. DeepSeek makes strong V4 claims in the release note, but many detailed charts are images. So the fair comparison is not "who wins every benchmark." It is "what does each provider officially claim, and what can we rely on?"
Official benchmark and capability claims
| Area | DeepSeek V4 | Qwen3-Max-Thinking |
|---|---|---|
| Agentic coding | DeepSeek says V4 Pro is open-source SOTA in agentic coding benchmarks | Qwen reports 75.3 on SWE Verified |
| Competitive coding | DeepSeek says V4 Pro beats current open models in coding | Qwen reports 85.9 on LiveCodeBench v6 |
| Science / reasoning | DeepSeek says V4 Pro beats current open models in Math, STEM, and coding | Qwen reports 87.4 on GPQA and 98.0 on HMMT Feb 25 |
| Agentic search | DeepSeek release focuses on agent integration and 1M context | Qwen reports 49.8 on HLE with tools |
| Cost-sensitive routing | V4 Flash is the clear first test | Qwen is stronger when the Alibaba/Qwen stack matters more than raw cost |
Source: DeepSeek V4 release note; Qwen3-Max-Thinking official blog
Do not overread provider benchmarks
These are provider claims. They are useful for direction, but they are not a substitute for testing your own prompts, codebase, latency limits, and cost profile.
Which model should you choose?
Match the model to the job.
Decision table
| Workload | Best first test | Why |
|---|---|---|
| Long-context document processing | DeepSeek V4 Flash | 1M context and low output pricing |
| High-volume agent worker calls | DeepSeek V4 Flash | Cheaper than Qwen on official listed prices |
| Harder open-weight reasoning experiments | DeepSeek V4 Pro | Flagship DeepSeek V4 model with open weights linked |
| Alibaba Cloud production stack | Qwen3-Max-Thinking | Native Model Studio support and tiered pricing docs |
| Tool-use reasoning inside Qwen | Qwen3-Max-Thinking | Qwen describes adaptive tool use and Model Studio lists tool calling |
| Cost stability after May 31 | Re-check before choosing | DeepSeek V4 Pro discount is temporary; V4 Flash pricing is more stable in the docs |
If I had to pick a default for most teams, I would start with DeepSeek V4 Flash. It is cheap enough to test broadly, it supports long context, and it keeps the migration path simple. I would test Qwen3-Max-Thinking when the app already depends on Alibaba Cloud or when Qwen's specific benchmark strengths line up with the workload.
Official sources checked
No old V3-only benchmark mirrors used.
- DeepSeek V4 Preview Release
- DeepSeek Models & Pricing
- DeepSeek API Change Log
- Qwen: Qwen3-Max-Thinking official blog
- Alibaba Cloud Model Studio supported models
- Alibaba Cloud Model Studio pricing
The bottom line
DeepSeek is the cost default. Qwen is the ecosystem pick.
The old headline was too broad. Chinese models are not automatically "beating GPT" at everything, and this page should not pretend one benchmark settles the market. The useful update is narrower and more practical: DeepSeek V4 is now live and should be tested first when you need cheap 1M-context API work. Qwen3-Max-Thinking is the better Qwen-side option when your deployment, tools, or evaluation target already live in the Alibaba/Qwen ecosystem.
For the DeepSeek-only details, read the DeepSeek V4 release and pricing guide. For broader closed-vs-open model choice, use the AI Model Picker. If you care most about monthly spend, compare your usage in the AI cost calculator.
Keep Reading
Need help choosing the right AI model?
Use our free tools to compare AI models by use case, cost, context length, and workflow fit.
Try the AI Model Picker