Qwen (Alibaba Tongyi Qianwen) Token Counter & Pricing (2026)

Qwen (Alibaba Tongyi Qianwen, 通义千问) is Alibaba's flagship LLM family, served through Bailian / Alibaba Cloud Model Studio. As of August 2026, the live API bills in RMB at ¥0.2–¥24 per million input tokens and ¥2–¥72 per million output tokens, depending on tier. The same weights are also reachable from outside China through Together AI, Fireworks AI, and OpenRouter, often at a different (sometimes lower, sometimes higher) rate than Alibaba's direct card. This page is the long-form breakdown of those numbers — what each Qwen model costs, how the Qwen tokenizer handles Chinese and English differently, and where overseas developers actually send their traffic.

If you want a pre-flight estimate before hitting the API, paste your text into the AI Token Calculator and pick a Qwen model from the dropdown. The rest of this page explains what that calculator is doing under the hood.

Where Qwen Stands in 2026

Three things separate Qwen from Western frontier providers in 2026.

First, it's the strongest Chinese tokenizer in production. Qwen ships a byte-level BPE tokenizer with a 152,064-token vocabulary, trained with deep coverage of Chinese, English, and code. That sounds like an implementation detail until you bill a Chinese-heavy workload: a prompt that costs 1,500 tokens on GPT-4o often costs 900–1,100 on Qwen. Same answer, 30–40% fewer tokens. We'll quantify this in the next section.

Second, it runs dual-track. Closed-API tiers (Qwen3.5-Flash, Qwen3.6-Flash, Qwen3.7-Plus, Qwen3.7-Max, Qwen3.8-Flash, Qwen3.8-Max-Prime, Qwen3-Max-Thinking) sit on one side. Apache-2.0 open weights (Qwen3-0.6B through Qwen3-235B-A22B, plus Qwen3-Coder and Qwen3-VL variants) sit on the other. Self-host the open weights on your own GPU and the per-token cost drops to zero. Call the hosted API and you pay per million tokens. The break-even threshold is somewhere around 50–100M tokens per month — below that, the API wins on capex; above that, self-hosting usually wins.

Third, the API is OpenAI-compatible. Base URL https://dashscope.aliyuncs.com/compatible-mode/v1, drop in an Alibaba Cloud API key from bailian.console.aliyun.com, set model to qwen-plus or qwen3-7-max, and most OpenAI SDK code works unchanged. That's why overseas tools like Cursor, Cline, and OpenCode can swap Qwen in without rewriting their client.

Note: I'm separating Qwen from GLM (Zhipu / 智谱) intentionally. They often get bundled together as "Chinese frontier models" but they bill differently, tokenize differently, and self-host on different stacks. Different problem.

Qwen Models & 2026 Pricing

All figures below come from help.aliyun.com/zh/model-studio/ and bailian.console.aliyun.com, verified 2026-08-28 to 2026-09-02. RMB→USD at 0.1389 (August 2026 mid-market). USD rates are Alibaba's published international/Singapore card; RMB rates are the China mainland endpoint.

A few things to notice.

Qwen3.5-Flash at ¥0.20/¥2 is the cheapest commercial frontier-tier model in this roundup. Not the cheapest total — DeepSeek and the smaller GLM tiers undercut it — but among the Qwen family on the paid API, this is the floor. Useful for classification, routing, extraction, anything where the answer is short.

Qwen3.8-Max-Prime at ¥24/¥72 is the most expensive. It's a premium-tier "优速" mode on top of the Max line: faster inference, higher priority, higher rate. Production agentic workloads where latency matters more than cost.

Tiering on Qwen3.7-Plus jumps inside a single request. Below 256K input tokens, you pay ¥2/¥8. Cross 256K and the rate jumps to ¥6/¥24 for the rest of the prompt. Long-context summarization at 800K tokens is three times the per-token cost of the same job at 200K.

Qwen3-Max-Thinking bills chain-of-thought as output, at the highest rate. Pass enable_thinking: true and the model emits a reasoning trace before its answer. That trace is billed at ¥43.20/MTok output — roughly 12× the standard Qwen3.6-Flash output rate. Reasoning is expensive. Turn it off when you don't need it.

Qwen Tokenizer Efficiency on Chinese

Here's the part that doesn't show up in the rate card but matters most for CJK workloads.

I ran the same mixed-language prompts through Qwen's tokenizer (Qwen/Qwen3-235B-A22B-Instruct-2507 from Hugging Face) and GPT-4o's tokenizer (o200k_base) on 2026-09-03. The result: Qwen wins on Chinese, ties on code, and loses slightly on pure English.

The Chinese numbers are the meaningful ones. If your workload is bilingual or CJK-heavy, the 30% token compression on the input alone cuts your bill by roughly the same percentage — without changing the model, the rate, or the answer.

Mechanically: Qwen's 152K vocab gives it dedicated tokens for common Chinese characters (single-byte or two-byte BPE merges), common code patterns (def, return, indentation markers), and high-frequency CJK identifiers. GPT-4o's tokenizer was trained primarily on English; Chinese characters fall back to UTF-8 byte sequences, which run 2–4 tokens per character on common glyphs.

Practical implication: for Chinese-document pipelines (contract review, customer support in Chinese, long-context summarization of Chinese text), Qwen is cheaper than GPT-4o at the token level, before you even compare the per-token rate. The two effects compound.

For code, the gap is small. Qwen's tokenizer has dedicated tokens for Python and JavaScript patterns, but the actual savings on real code (which is already 60–80% ASCII) are 2–5%, not the 30% you'll see on Chinese prose.

For pure English prose, GPT-4o wins by a hair. Qwen's larger vocab has slightly more overhead on sequences where every word already has a dedicated token.

Overseas Hosting Options

Alibaba's direct API works from outside China — sign up at bailian.console.aliyun.com, get a Singapore endpoint, drop in your credit card. But three alternatives route around the Alibaba account entirely.

Together AI (together.ai). Hosts the Qwen3 open-weights line as managed inference. You get an OpenAI-style API key, point your client at api.together.xyz/v1, set model to Qwen/Qwen3-235B-A22B-Instruct-2507 (or the variant you want). Together bills per million tokens, often with different rates than Alibaba's direct card. Together is generally cheaper for high-volume batch jobs because they pass through the open-weight cost structure (no closed-API margin). For Qwen3-235B specifically, expect Together to be 20–40% below Alibaba's direct ¥-based card.

Fireworks AI (fireworks.ai). Same model lineup, different optimization stack. Fireworks focuses on low-latency serving — their Qwen3 deployments tend to have faster TTFT (time to first token) than Together, with slightly higher per-token cost. For interactive chat and coding assistants where latency is the bottleneck, Fireworks is usually the right pick.

OpenRouter (openrouter.ai). Aggregates both, plus a dozen other Qwen hosts. OpenRouter adds Qwen to a single OpenAI-style key and lets you route by price, latency, or context length on a per-request basis. Useful when you don't want to commit to one host — you can switch providers via model-string swap (qwen/qwen-3-235b-a22b on OpenRouter routes to whichever host currently has the best latency for your region). OpenRouter charges a small routing fee on top of the host's rate.

One thing worth knowing: rates on these routers change weekly. What Together charges for Qwen3-235B today may not be what they charge next month. Pin your budget model to the host's published rate card at the time you commit, not at the time you read this article.

Real Bill Examples

Four worked scenarios using the rate table above. All in USD at Alibaba's direct international rate.

Example 1: Chinese-language customer support chatbot on Qwen3.7-Plus. 1,500-token system prompt + 200-token user message (Chinese), 350-token Chinese reply, 40,000 conversations a month. - Input: 1,700 × 40,000 = 68M tokens × $0.28 / 1M = $19.04 - Output: 350 × 40,000 = 14M tokens × $1.11 / 1M = $15.54 - Total: $34.58/month - On Qwen3.5-Flash (same workload, shorter context): $1.89 + $3.89 = $5.78/month. - On Qwen3.8-Flash: $7.55 + $5.25 = $12.80/month. - The token-compression advantage on Chinese is already baked in; the rate card advantage is the second lever.

Example 2: Bilingual contract summarization pipeline. 8,000-token mixed CN/EN legal contract, 1,000-token summary, 5,000 contracts a month on Qwen3.7-Plus. - Input: 40M tokens × $0.28 = $11.20 - Output: 5M tokens × $1.11 = $5.55 - Total: $16.75/month - Same workload on Qwen3.7-Max: $66.80 + $27.75 = $94.55/month. Five times the cost, and unless your summarization quality genuinely demands it, the Plus tier is the right pick.

Example 3: Agentic coding on Qwen3-Max-Thinking. A typical Cline / Cursor / Claude Code day burns 5–10M tokens with cache reads dominating. With thinking enabled, output tokens run 3–5× the standard rate. A 7M-token day split roughly 60/40 input/output on Qwen3-Max-Thinking at $1.20 input / $6.00 output: - Input: 4.2M × $1.20 / 1M = $5.04 - Output: 2.8M × $6.00 / 1M = $16.80 - Daily: $21.84. Monthly (20 working days): $436.80. - Same workload on Qwen3.6-Flash without thinking: $0.70 + $2.80 = $3.50/day, $70/month. Six times cheaper. The question is whether the thinking mode catches bugs the Flash tier misses — for most coding workloads it does, but the cost is real.

Example 4: High-volume routing/classification on Qwen3.5-Flash. 1M short requests a month, average 200-token input + 50-token classification output. - Input: 200M tokens × $0.03 / 1M = $5.56 - Output: 50M tokens × $0.28 / 1M = $13.89 - Total: $19.45/month for a million requests. - At ¥0.20/¥2 on the China endpoint, the same million requests: ¥40 + ¥100 = ¥140 ≈ $19.45. Roughly the same — RMB→USD conversion is doing the work.

The pattern across all four: Qwen3.5-Flash and Qwen3.6-Flash handle anything where the answer is short or the volume is high. Qwen3.7-Plus is the steady middle. Qwen3.7-Max and the thinking/Prime tiers are the right pick only when you can quantify the quality gain and bill for it.

Frequently Asked Questions

What's the cheapest Qwen model on the paid API? Qwen3.5-Flash at ¥0.20 input / ¥2.00 output per million tokens (≈$0.03 / $0.28). Confirmed against Alibaba Model Studio on 2026-08-28. Below that tier, you go to self-hosted open weights (Qwen3-0.6B / Qwen3-4B / Qwen3-30B-A3B) and pay only your GPU cost.

Why is Qwen cheaper on Chinese prompts than GPT-4o? Two compounding effects. First, Qwen's tokenizer compresses Chinese characters into roughly 1.2 tokens each vs GPT-4o's 1.8–2.5, a 30–40% token-count reduction. Second, Qwen's per-token rates on Plus and Flash are 3–10× below GPT-4o's, even before that compression. A bilingual workload that costs $100 on GPT-4o typically costs $15–25 on Qwen3.7-Plus.

Can I call Qwen from outside China without an Alibaba Cloud account? Yes. Three routes. Together AI hosts Qwen3 open weights under a standard API key, often 20–40% below Alibaba's direct rate. Fireworks AI does the same with lower latency, slightly higher rate. OpenRouter aggregates both and lets you swap providers per request. All three accept OpenAI-style SDK code with a base-URL change.

Does Qwen bill reasoning tokens separately? On Qwen3-Max-Thinking, no — reasoning tokens are billed as output tokens at the model's standard output rate (¥43.20/MTok, $6.00/MTok). On Qwen3.7-Plus and Qwen3.7-Max, reasoning is opt-in via enable_thinking: true and also bills as output. On the Flash tiers, reasoning is not exposed. If you turn thinking on, expect your output bill to climb 3–5×.

Should I self-host Qwen3 open weights or call the API? Roughly: above 50–100M tokens per month, self-hosting on rented H100s or 4090s usually wins on cost. Below that, the API wins because you avoid capex and ops. The exact break-even depends on your token-mix (input-heavy workloads favor self-hosting because prompt caching on the API narrows the gap), your region (cheaper electricity in some markets), and how much your engineering time is worth. As a rule of thumb, if you're a 3-person team spending more than $2,000/month on the Qwen API, run the self-hosting math once.

Sources

1. Alibaba Cloud Model Studio — Qwen Model Pricing. help.aliyun.com/zh/model-studio/. Verified 2026-08-28 to 2026-09-02.
2. Qwen3 open-weights repository and tokenizer configs. github.com/QwenLM/Qwen3 and huggingface.co/Qwen. Verified 2026-09-03.
3. Alibaba Cloud Bailian Console — API key management and free-tier quota. bailian.console.aliyun.com. Retrieved 2026-09-03.
4. Qwen3-235B-A22B-Instruct-2507 tokenizer card. huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507. Verified 2026-09-03.
5. Together AI — Qwen model catalog and pricing. api.together.ai/models. Retrieved 2026-09-03.
6. Fireworks AI — Qwen model catalog and pricing. fireworks.ai/models. Retrieved 2026-09-03.