If you have been following the AI leaderboards lately, you have likely noticed a name that keeps trading blows with the field’s proprietary leaders: Qwen.

Developed by Alibaba Cloud, the Qwen model family has kept expanding at a rapid pace — the current lineup spans Qwen3, Qwen3.5, Qwen3.6, and the flagship Qwen3.8-Max, alongside dedicated coding and vision variants. Qwen remains widely considered a leader in coding and math among open-weight models, and it continues to close the gap with frontier proprietary systems while staying dramatically cheaper to run.

Since Alibaba releases the weights for most of these models, you aren’t forced to use a single proprietary API. This has created a competitive market where providers race to offer the lowest price. This guide cuts through the noise to give you the definitive pricing strategy for Qwen.

Executive Summary: The Qwen Pricing Cheatsheet

If you just want the quick answer on where to go to save the most money, here is your cheat sheet.

Best For	Provider Recommendation	Why?
Lowest Price & Best Variety	DeepInfra	Near-at-cost pricing across the widest range of Qwen models, including Coder, Vision, and MoE variants.
Proprietary Flagship (Qwen3-Max)	DeepInfra or Alibaba Cloud	DeepInfra now hosts Qwen3-Max and Qwen3-Max-Thinking directly, at the same entry-tier rate Alibaba charges first-party. Go to Alibaba only for Chinese-region hosting or batch-API discounts.
Easiest to Start	OpenRouter	User-friendly aggregator with strong documentation, though typically priced slightly above DeepInfra.
Developers using RAG	DeepInfra	Publishes explicit cached-input rates on several Qwen models, which creates real savings for document-heavy apps.

Understanding the “Pay-as-You-Go” Token Model

Before looking at the price tags, it’s crucial to understand what you’re paying for. AI providers charge per token.

Think of a token as a piece of a word. Roughly, 1,000 tokens equals about 750 words.

• Input Tokens: The text you send to the model (your prompt, documents, chat history).
• Output Tokens: The text the AI generates in response.

The “Chat History” Trap: For a chatbot to “remember” a conversation, you must re-send the entire chat history with every new message. This means your Input Token usage grows with every turn, making low input prices the most critical factor for cost savings.

Provider Spotlight: DeepInfra Pricing

DeepInfra has established itself as the “power user’s choice” for Qwen. Because they run on bare-metal infrastructure without the massive overhead of a general-purpose cloud, they offer rates that are often 50-80% cheaper than major competitors.

You can view their full list of Qwen models here: DeepInfra Qwen Models.

Here is the current pricing breakdown for the most popular Qwen options on their platform:

Model Name	Best Use Case	Context Window	Input Price (per 1M)	Output Price (per 1M)
Qwen3-Max-Thinking	Overall Best. DeepInfra’s flagship reasoning model, with adaptive tool-use and test-time scaling.	256K	$1.20 ($0.24 cached)	$6.00
Qwen3-Coder-480B-A35B-Instruct-Turbo	Coding. DeepInfra’s most agentic code model, competitive with Claude Sonnet on coding benchmarks.	262K	$0.30 ($0.10 cached)	$1.00
Qwen3-VL-235B-A22B-Instruct	Vision. Reasons over images, charts, screenshots, and PDFs alongside text.	256K	$0.20 ($0.11 cached)	$0.88
Qwen3.5-27B	Mid-Range. The current-generation “Goldilocks” model — sharper than small models, lighter than the flagship.	256K	$0.26	$2.60
Qwen3.5-9B	Speed & Cost. Fast, cheap, long context. Ideal for classification, summarization, or simple bots.	256K	$0.10	$0.15
Qwen3.5-35B-A3B	Mixture of Experts (MoE). Efficient sparse-activation model with native cache support.	256K	$0.14 ($0.05 cached)	$1.00

Note: Prices are per 1 million tokens. Cached-input rates apply where the provider offers explicit prompt caching.

Why this matters: At $1.20 per million input tokens, Qwen3-Max-Thinking is roughly 4x cheaper than OpenAI’s current flagship, GPT-5.6 Sol ($5.00/1M input), while remaining competitive on reasoning and agentic benchmarks.

Official Source: Alibaba Cloud Pricing

Alibaba Cloud is the creator of Qwen. While their platform is excellent, it is generally more complex to navigate than Western API wrappers. However, you must use them if you need Qwen-Max.

Alibaba Cloud Model Studio Pricing

Model	Type	Input Price (per 1M)	Output Price (per 1M)
Qwen3.8-Max	Proprietary Flagship	$2.00	$6.00
Qwen3.7-Plus	Balanced	$0.40 (0–256K tier; rises to $1.20 above 256K)	$1.60 (rises to $4.80 above 256K)
Qwen3.8-Flash	Fast & Cheap	$0.15	$0.47

Note: Prices shown are Alibaba Cloud Model Studio’s standard International (Singapore) rates, verified directly against Alibaba’s official pricing documentation (last updated September 1, 2026). Several models use tiered pricing that increases with prompt length — the rates above are entry-tier.

Alibaba Cloud is the creator of Qwen, and its platform is the first-party source for every model release, often ahead of third-party providers. It’s also the place to go for Chinese-region hosting, batch-API discounts, or a model version DeepInfra hasn’t mirrored yet.

That said, the old advice to “go to Alibaba if you need the proprietary flagship” no longer fully holds: DeepInfra now hosts Qwen3-Max and Qwen3-Max-Thinking directly, at the same $1.20/$6.00 entry-tier rate Alibaba charges first-party. The real gap is on the cheap end — Qwen3.8-Max’s output price ($6.00) is roughly 40x the output price of DeepInfra’s cheapest current Qwen model, Qwen3.5-9B ($0.15).

The Hidden Cost-Saver: Context Caching

This is the secret weapon for building cheap AI apps.

Imagine you have a 50-page employee handbook. You want employees to be able to ask questions about it. Without caching, you have to pay to send that 50-page handbook (approx. 25k tokens) to the model every single time a user asks a question.

DeepInfra publishes explicit cached-input rates on several Qwen models. A few examples:

• Qwen3-Max-Thinking: $1.20 → $0.24 per 1M cached tokens (an 80% discount)
• Qwen3-Coder-480B-A35B-Instruct-Turbo: $0.30 → $0.10 (a 67% discount)
• Qwen3.5-35B-A3B: $0.14 → $0.05 (a 64% discount)

On Alibaba Cloud’s own platform, cache hits are billed at a flat 10% of the standard input rate (a 90% discount) across supported models.

If you are building a “Chat with PDF” tool or a bot with a long system prompt, caching can meaningfully cut your bill. DeepInfra supports this feature across most of its current Qwen lineup.

Real-World Cost Scenarios

Let’s translate these abstract numbers into actual monthly bills.

Scenario A: The Customer Support Bot

• Volume: 5,000 tickets/month
• Complexity: 5 turns per ticket
• Model: Qwen3.5-9B (via DeepInfra)
• Total Tokens: ~25 million input, ~1 million output

EsEstimated Cost:

• Input: 25M × $0.10 = $2.50
• Output: 1M × $0.15 = $0.15
• Total Monthly Bill: ~$2.65

(Compare this to roughly $155/month on GPT-5.6 Sol at the same volume.)

Scenario B: The Coding Assistant

• Task: A dev team using an AI coding assistant inside VS Code
• Model: Qwen3-Coder-480B-A35B-Instruct-Turbo
• Volume: Heavy daily usage (code completion + refactoring)
• Total Tokens: ~100 million input, ~5 million output

Estimated Cost:

• Input: 100M × $0.30 = $30.00
• Output: 5M × $1.00 = $5.00
• Total Monthly Bill: ~$35.00

Conclusion: Which Path to Choose?

For 95% of developers and businesses, the days of paying expensive premiums for top-tier AI are over. Qwen 2.5 72B offers For most developers and businesses, the days of paying steep premiums for top-tier AI are over. Qwen’s current flagship models rival the world’s best proprietary systems at a fraction of the price.

• Go with DeepInfra for the standard Qwen experience — including the proprietary Qwen3-Max line — at competitive market rates with excellent API compatibility.
• Go with Alibaba Cloud directly if you need Chinese-region data compliance, batch-API pricing, or a model release DeepInfra hasn’t added yet.

By choosing the right model and provider, you can build production-grade AI applications for the price of a few lattes a month.