Contents
- Verdict
- API pricing
- The tiered pricing trap, and how to design around it
- What Qwen costs on two real automations
- Agent reliability and the shape of the pricing
- Open models that actually fit on hardware you can buy
- The verdict for a small business
- How to access
- Privacy and residency
- What we build
- FAQ
- Price sources
Qwen is the Chinese provider that is easiest to justify to a cautious stakeholder. Alibaba Cloud Model Studio publishes an explicit pledge never to use your data for model training, runs international deployments in Singapore, Hong Kong, Germany, the US and Japan, gives cache hits at 10 percent of input, batch inference at 50 percent, and a free quota of 1M tokens per model valid for 90 days. Pricing is competitive and in places genuinely cheap: qwen-flash starts at $0.05 in and $0.40 out per 1M tokens and qwen3.5-flash at a flat $0.10 and $0.40 as of August 2026. Two design details do the damage if you ignore them. Input is priced by length band, so long prompts cost several times more per token. And batch pricing and cache discounts are mutually exclusive, so you cannot stack the two cheapest levers.
Best for
Businesses that want cheap Chinese-lab capability with a clear no-training commitment, and anyone who wants genuinely single-GPU-hostable open models with an Apache 2.0 licence.
Not for
Long-context agent loops on the metered API, where the tiered pricing escalates sharply once prompts grow past a few hundred thousand tokens.
Alibaba Qwen API pricing per 1M tokens
Alibaba prices most Qwen models by input length band, so the same model costs several times more per token on a long prompt than a short one. That is the single most important thing to understand before you build an agent on it, and it is why qwen3.5-flash at a flat $0.10 and $0.40 across the whole 0 to 1M range is often the sensible pick despite not being the cheapest headline. Cache hits bill at 10 percent of the standard input price while explicit cache creation costs 125 percent, so a cache needs at least a couple of hits to pay for itself. Batch inference is 50 percent of the real-time price for both input and output, and Alibaba documents that batch pricing and the cache discount are mutually exclusive: you get one or the other, not both. Model Studio offers a free quota of 1M tokens per model in the international deployment scope, valid 90 days. Prices below are international-region list prices as of August 2026, read from a pricing document last updated 15 July 2026.
| Model | Context | Input / 1M | Output / 1M |
|---|---|---|---|
| qwen3.7-max (flagship)Cache hit 10 percent of input; batch 50 percent; 1M-token free quota for 90 days | Single 0 to 1M band | $2.50 | $7.50 |
| qwen3-maxRises to $2.40 then $3.00 input, and $12.00 then $15.00 output, by band | Banded 0-32K / 32-128K / 128-256K | $1.20 | $6.00 |
| qwen3.7-plusRises to $1.20 input and $4.80 output above 256K | Banded 0-256K / 256K-1M | $0.40 | $1.60 |
| qwen3.6-flashRises to $1.00 input and $4.00 output above 256K | Banded 0-256K / 256K-1M | $0.25 | $1.50 |
| qwen3.5-flashFlat rate with no length banding, which makes it the predictable choice for agent loops | Single 0 to 1M band | $0.10 | $0.40 |
| qwen-flash (cheapest)Rises to $0.25 input and $2.00 output above 256K | Banded 0-256K / 256K-1M | $0.05 | $0.40 |
| qwen-long-latestNo free quota; aimed at long-document work | Not published in the pricing table | $0.072 | $0.287 |
| Qwen3.8-MaxStill absent from Alibaba's own international pricing documentation when we checked in August 2026, so we are not quoting third-party figures for it | Not published | Not published | Not published |
List prices as of August 2026, taken from the provider's published pricing page on August 5, 2026. Model prices move fast - check the source links before you budget.
The tiered pricing trap, and how to design around it
Most providers charge one input rate whatever the prompt length. Alibaba charges by band, and the bands rise sharply. On qwen3.6-flash, input goes from $0.25 per 1M below 256K tokens to $1.00 above it, and output from $1.50 to $4.00. On qwen3.7-plus it is $0.40 rising to $1.20 on input and $1.60 rising to $4.80 on output. On qwen3-max there are three bands: $1.20, $2.40 and $3.00 on input and $6.00, $12.00 and $15.00 on output. Same model, up to two and a half times the unit price, decided entirely by how long your prompt got.
The exception is worth knowing because it solves the problem outright. qwen3.5-flash is priced at a flat $0.10 in and $0.40 out across the whole 0 to 1M range with no banding at all, and qwen3.7-max sits in a single 0 to 1M band at $2.50 and $7.50. If you are building an agent whose context you cannot bound tightly, those flat-rate models are worth more than a lower headline rate on a banded one, because a forecast you can trust beats a price you cannot.
For ordinary request-response automation this barely matters, because a classification or extraction prompt is a few thousand tokens and sits comfortably in the cheapest band. For agent loops it matters enormously, because an agent accumulates context with every turn. A long-running agent can start in the cheapest band and finish in the most expensive one without anyone changing a line of code, and the bill will not look like your estimate.
The design response is the same one that is good practice anyway: keep working context bounded. Summarise or compact the conversation history at intervals, drop tool outputs that are no longer relevant, and carry forward a structured state object rather than the full transcript. On Qwen that discipline is not just tidy engineering, it directly controls the per-token rate you pay.
One correction to a claim you will see elsewhere, including in earlier versions of this page: those discounts do not stack. Alibaba's billing documentation states that batch pricing and the context cache discount are mutually exclusive. You choose one. For a chatty agent with a heavy system prompt the cache is usually worth more; for an overnight bulk job with little prompt repetition, batch wins. Working out which applies to your workload takes ten minutes and is worth doing before you design around the wrong one.
What Qwen costs on two real automations
We price every provider on this site against the same two workloads. Workload A is an inbox agent: 1,500 runs a month at about 3,000 input tokens each, of which 2,400 are a stable prefix, plus 400 output tokens. Workload B is document extraction: 4,000 documents a month at 6,000 input and 300 output tokens each. Both sit comfortably inside the cheapest length band on every model, which is the good case for Qwen.
On qwen3.6-flash at $0.25 in and $1.50 out, Workload A costs about $2.03 a month, falling to roughly $1.22 if the stable prefix hits cache at 10 percent of input. Workload B costs about $7.80, or about $3.90 through batch inference at half price. Remember that those are alternatives rather than additions: you cannot batch and cache the same tokens.
On qwen-flash at $0.05 and $0.40 the same workloads are about $0.47 and $1.68, with batch taking B to about $0.84. On the flat-rate qwen3.5-flash at $0.10 and $0.40 they are about $0.69 and $2.88. On qwen3.7-plus at $0.40 and $1.60, about $2.76 and $11.52. On the qwen3.7-max flagship at $2.50 and $7.50, about $15.75 and $69.00, or $34.50 batched.
Two things distort comparisons here and both are worth guarding against. First, the free quota: 1M tokens per model for 90 days in the international scope means your first month of testing is genuinely free, which is great for evaluation and useless for forecasting. Second, region. Alibaba's own documentation shows Beijing pricing at roughly 70 percent below international for the same model, so any comparison table quoting Qwen as dramatically cheaper than everyone else is probably quoting Beijing rates at a business that will be billed international ones. The Hong Kong, Frankfurt and US regions do carry documented time-based promotions of 80 percent off at night and 50 percent off during the day, which is a real discount but a promotional one to re-check rather than build a budget on.
Agent reliability and the shape of the pricing
Qwen's specific reliability risk is not the model, it is the billing shape interacting with how agents behave. An agent accumulates context every turn. On a banded model, an agent that starts a run at 40K tokens and finishes at 300K has quietly moved into a band charging four times as much per token, and it did so without any code change or alert. Nothing about that is a model defect, but it will make a monthly bill look nothing like the estimate.
The mitigation is the same discipline the published research recommends for accuracy. LOCA-bench, released in early 2026, grows agent context under controlled conditions while holding the task fixed and reports that most models including frontier and open-source ones degrade sharply as context grows. AgentCE-Bench, from April 2026, found that a 10 percent tool-call failure rate visibly degrades performance at every horizon tested. Compacting context between steps, dropping stale tool results, carrying a structured state object rather than a full transcript, and adding retry and backoff with idempotency on writes are the answers to both the accuracy problem and the banding problem at once.
If you cannot bound context tightly, pick a flat-rate model. qwen3.5-flash at $0.10 and $0.40 across the whole range removes the banding risk entirely for a modest premium over qwen-flash's short-prompt rate. That is a rare case where the right answer is not the cheapest line on the price list, and it is the recommendation we would make for most agent builds on this provider.
As always, we publish no benchmark scores. Run twenty of your own real cases through qwen3.6-flash and whatever you use today, grade them with someone who knows the right answers, and let the free quota pay for the experiment.
Open models that actually fit on hardware you can buy
Most open-weight flagships are open in the licensing sense but not in the practical sense: you can download them, and then you discover you need a multi-node cluster to serve them. Qwen is the notable exception. Alongside the closed Max flagships, Alibaba publishes Apache 2.0 models in sizes designed to run on real hardware - a mixture-of-experts model with only a few billion active parameters per token, and a dense model in the tens of billions, both with very large native context.
Apache 2.0 is worth dwelling on. It is a genuine OSI-approved open-source licence with no revenue thresholds, no branding requirements and no acceptable-use riders. That is a cleaner legal position than Meta's Llama Community Licence, which is often described as open source but is not one, and it removes the need for a legal review before you deploy.
For a small business the practical implication is that a self-hosted automation is actually achievable. One or two workstation-class GPUs can serve a capable model for internal use, with no per-token billing and no data leaving your network. That is the honest answer to the data-residency question that no API provider can offer, and Qwen is the most realistic route to it.
Do the sums before committing. A GPU server plus the engineering time to keep it patched, monitored and available will usually exceed a few hundred dollars a month of API spend, so self-hosting on cost grounds alone rarely pays for a small business. Self-hosting because the data legally cannot leave your premises is a different and much better reason.
The verdict for a small business
Qwen occupies a useful position: cheaper than the US majors, with a data-handling commitment that reads much like theirs, on infrastructure contracted through a large commercial cloud rather than a startup. For a business that wants the cost advantage of the Chinese labs but needs to be able to answer a procurement question, it is the easiest of the four to defend.
It is not the cheapest. Z.ai's FlashX tier and DeepSeek's cache-hit pricing both go lower. What Qwen offers instead is the combination of low cost with a clear no-training pledge and a familiar contracting relationship, which for many businesses is worth more than the last few cents per million tokens.
The volatility caveat applies with particular force here. Qwen3.8-Max circulates widely but was still absent from Alibaba's own international pricing documentation when we re-checked in August 2026, against a document last updated on 15 July 2026, with qwen3.7-max still the newest Max-tier entry. We are not going to print aggregator numbers for a model the vendor has not priced publicly. If you want to use it, get the rate from Alibaba before you budget.
Where we would use it in a build: short-prompt high-volume work on the metered API, overnight batch processing at half price, and self-hosted deployment of the small open models where data cannot leave the network. Where we would look elsewhere: long-context agent loops, where the tiered pricing works against you and a flat-rate provider will be cheaper and easier to forecast.
Strengths for agent workloads
- An explicit published statement that customer data will never be used for model training, the clearest such pledge among the Chinese providers
- International deployments in Singapore, Hong Kong, Germany, the US and Japan under Alibaba Cloud's ordinary cloud contracting, which is familiar territory for procurement
- qwen3.5-flash at a flat $0.10 in and $0.40 out with no length banding, which makes agent costs forecastable
- qwen-flash at $0.05 per 1M input on short prompts, among the cheapest rates published anywhere
- Apache 2.0 open models in genuinely small sizes, including a mixture-of-experts model with a few billion active parameters that runs on one or two GPUs
- Cache hits at 10 percent of input price and batch inference at 50 percent
- A 1M-token free quota per model, valid 90 days, so you can evaluate before spending anything
- Time-based promotions on the Hong Kong, Frankfurt and US regions, documented as 80 percent off at night and 50 percent off during the day
Weaknesses
- Tiered input pricing rises steeply with prompt length on most models, which is exactly the wrong shape for long-running agent loops
- Batch pricing and the cache-hit discount are mutually exclusive, so the two cheapest levers cannot be stacked the way they can on OpenAI or Anthropic
- Explicit cache creation costs 125 percent of the standard input price, so a cache that is never hit twice loses money
- Context windows are not stated in the pricing document, only input-length billing bands, so capacity planning needs the model pages
- The flagship Max models are closed weights, so the open-model story only covers the smaller tiers
- Retention periods for API inputs and outputs are not clearly published beyond telemetry, and the physical storage location within a region is not specified
- Qwen3.8-Max was still absent from Alibaba's own international pricing documentation in August 2026 despite circulating elsewhere
- Beijing pricing runs roughly 70 percent below international pricing for the same model, so comparing the wrong region flatters Qwen enormously
How to access Alibaba Qwen
- Alibaba Cloud Model Studio in the international deployment scope, which covers Singapore, Hong Kong, Germany, the US and Japan. This is the route for businesses outside China and it comes under ordinary Alibaba Cloud contracting, with a 1M-token free quota per model for 90 days.
- The separate Beijing endpoint, which is a distinct service with much cheaper pricing - roughly 70 percent below international rates on the same model - and a different scope. Do not budget from those figures if you will use the international region, and be sceptical of any comparison table that quotes them without saying so.
- OpenRouter, which carries a very large number of Qwen models including the newest flagship from day one, useful for testing without opening a cloud account.
- Self-hosting the Apache 2.0 open models. This is where Qwen is genuinely distinctive: the smaller mixture-of-experts and dense models are sized to run on one or two consumer or workstation GPUs, unlike most open flagships that need a data centre.
- Local coding agents using the open coding models, which is a practical way to keep source code entirely on your own machines.
Data privacy and residency
- Alibaba Cloud's published privacy notice for Model Studio states it will never use your data for model training. That is the most explicit no-training commitment of the Chinese providers in this comparison and matches the default position of the US majors.
- The international deployment scope covers Singapore, Hong Kong, Germany, the US and Japan, contractually separate from the Beijing endpoint. A German region in particular is a materially better residency answer for an EU business than anything else in this section of the directory offers. As with the other providers here, confirm which region and endpoint your integration actually uses.
- Retention is less clear than the training commitment. Beyond a stated 30-day window for telemetry, we could not find a published retention period for API inputs and outputs, and the physical storage location within the Singapore region is not specified. If either matters to you, ask before you build.
- Alibaba Cloud is a large commercial cloud provider with standard enterprise contracting, data processing agreements and compliance certifications available. In procurement terms that usually makes it easier to approve than a smaller AI-only vendor, whatever the flag on the headquarters.
- As with Z.ai, some organisations will have client, insurer or contractual constraints on using a China-headquartered provider regardless of the terms and the server location. That is a legitimate commercial constraint and separate from any technical assessment.
What we build on Alibaba Qwen
- Short-prompt, high-volume automation - classification, extraction, tagging - where the cheapest input band applies
- Batch document processing at 50 percent off for backlogs that can run overnight
- Self-hosted automation on the Apache 2.0 open models, which are small enough for a single well-specified server
- Local coding agents that keep source code on your own machines using the open coding models
- Cheap first-pass triage in a two-model stack, escalating only ambiguous cases to a flagship
Model choice is a line item, not a strategy. If you want the workflow designed, built and handed over with a plain-language report of what it ran and what it saved, start with a free AI automation audit or see the full done-for-you automation services.
Alibaba Qwen FAQ
How much does the Qwen API cost?
On the international deployment scope as of August 2026: qwen-flash from $0.05 per 1M input and $0.40 output, qwen3.5-flash at a flat $0.10 and $0.40, qwen3.6-flash from $0.25 and $1.50, qwen3.7-plus from $0.40 and $1.60, qwen3-max from $1.20 and $6.00, and qwen3.7-max at $2.50 and $7.50. Most of those rise by input-length band. Cache hits cost 10 percent of input, cache creation 125 percent, and batch inference 50 percent, with batch and cache mutually exclusive.
What would a real Qwen automation cost per month?
A 1,500-run inbox agent at about 3,000 input and 400 output tokens per run costs roughly $2.03 a month on qwen3.6-flash, or about $1.22 with cache hits on the system prompt. A 4,000-document extraction job at 6,000 input and 300 output tokens each is about $7.80, or about $3.90 batched. On qwen-flash the same two are roughly $0.47 and $1.68. Your first 1M tokens per model are free for 90 days, which covers the evaluation but not the forecast.
Can I combine Qwen's batch discount with prompt caching?
No. Alibaba's billing documentation states that batch pricing and the context cache discount are mutually exclusive, so you pick one. For a chatty agent resending a heavy system prompt every turn, caching is usually worth more. For an overnight bulk job with little prompt repetition, batch wins. Note too that explicit cache creation costs 125 percent of the standard input rate, so a cache that is only hit once loses money.
Which Qwen region should I use?
The international deployment scope covers Singapore, Hong Kong, Germany, the US and Japan, and a German region is a materially better residency answer for an EU business than most alternatives in this part of the directory. Beijing is a separate service priced roughly 70 percent lower, so never budget from Beijing figures for an international deployment. Hong Kong, Frankfurt and US regions carry documented promotional discounts of 80 percent off at night and 50 percent by day.
Does Alibaba train on Qwen API data?
Alibaba Cloud's published privacy notice for Model Studio states it will never use your data for model training. That is the clearest no-training commitment among the Chinese providers in this comparison and matches the default position of the US majors. Retention periods for API content are less clearly published, so ask if that matters to you.
Why does Qwen pricing change with prompt length?
Alibaba prices most models by input-length band. On qwen3.6-flash, input goes from $0.25 to $1.00 per 1M and output from $1.50 to $4.00 once a prompt passes 256K tokens. qwen3-max has three bands topping out at $3.00 input and $15.00 output. It barely affects ordinary request-response automation but it hits agent loops hard, since agents accumulate context. Either compact context aggressively, or use qwen3.5-flash, which is flat-rated across the whole range.
Is Qwen3.8-Max available and what does it cost?
We could not confirm either from Alibaba's own documentation. Qwen3.8-Max was still absent from the international pricing document when we checked in August 2026, against a page last updated 15 July 2026 in which qwen3.7-max is the newest Max-tier model. Figures for it circulate on aggregator sites and we are not going to repeat them as vendor pricing. Ask Alibaba for the rate before budgeting on it.
Can I self-host Qwen?
The smaller models, yes, and this is Qwen's standout feature. Alibaba publishes Apache 2.0 models sized to run on one or two workstation GPUs, including a mixture-of-experts model with only a few billion active parameters. The flagship Max models are closed weights. Apache 2.0 is a genuine open-source licence with no revenue thresholds or branding conditions.
Is the Qwen international endpoint different from the China one?
Yes, materially. The Singapore international region and the Beijing endpoint are separate services with different pricing and scope. The China endpoint is considerably cheaper, so do not budget from those figures if you will be using the international region. Confirm which endpoint your SDK defaults to before you go live.
Qwen or DeepSeek for business automation?
DeepSeek is cheaper, particularly on cache hits, and has MIT-licensed weights on both tiers. Qwen has the better data position for a business - an explicit no-training pledge, a Singapore endpoint and ordinary cloud contracting - plus open models small enough to self-host on hardware you can actually buy. For non-sensitive volume work, DeepSeek. For anything you need to justify to a client, Qwen.
