Qwen API Pricing in 2026: Every Tier, Kimi, GLM and MiniMax Compared

Alibaba Cloud Model Studio bills Qwen API calls pay-as-you-go by default: you pay for input tokens and output tokens separately, at a unit price set per model. Most English-language buyers start in the Singapore region, because that is the only region where new accounts get a free quota. The table below copies the standard prices from Alibaba's alibabacloud.com/help/en/model-studio/model-pricing page for the current text models.

Thinking mode. For qwen3.8-max and qwen3.7-max one price covers both thinking and non-thinking output. For qwen3.7-plus the non-thinking and thinking output prices are the same ($1.6 in the lowest tier). The older qwen-plus alias is the exception: $1.2 output without thinking but $4 with thinking, so check which mode your client enables.

Promotions. Alibaba marks qwen3.7-plus as a limited-time 20% off list price in Singapore and US scope, and some Global-scope models carry night-time discounts (22:00 to 08:00 UTC+8). The page states it lists standard prices and points to the Model Studio console for current promotions, so treat any discount as temporary.

Older aliases still bill. qwen-max (International) is $1.6 / $6.4, qwen-plus is $0.4 / $1.2 non-thinking, and qwen-flash is $0.05 / $0.4 up to 256K. If your code still calls an alias, the alias price applies, not the newest model's.

Tiers are decided per request, not per month. Alibaba's rule is that the total input tokens in a single request pick the tier, and every token in that request is billed at that tier's price. A 300K-token request to qwen3.7-plus is billed entirely at $1.2 / $4.8, not partly at $0.4. For a pipeline that classifies short posts, keeping each request small keeps you in the lowest tier.

Region and scope. Each region has its own endpoint, its own API keys and its own model list, and keys cannot be used across regions. The same model ID carries different prices by scope:

Prices are input / output per 1M tokens, as Alibaba prints them in USD. The Global scope is selectable from US (Virginia), Hong Kong and Germany; data residency and latency differ, so pick the scope your compliance team accepts before you pick the price.

Batch and cache. Where a model supports batch calls, both input and output are billed at 50% of the real-time price. Where it supports context caching, cache hits are billed at 10% of the input price and explicit cache creation at 125%. Alibaba states the two discounts cannot apply to the same call. A nightly job that re-reads the same long system prompt benefits from caching; an overnight bulk classification benefits from batch.

The steps below follow Alibaba's "Obtain an API key" and "Free quota for new users" pages for Model Studio.

Step 1: Create and complete the Alibaba Cloud account. Alibaba now requires completed account information before Model Studio can be activated. Then open Model Studio in the Singapore region and accept the service agreement; that activation grants the free quota automatically (it can take up to two hours to show).

Step 2: Create the key. On the Model Studio API Key page, select the region in the upper-right corner and click Create API Key. Choose the default workspace, optionally add a description, and pick All permissions or Custom (IP allowlist and model list). Copy the key immediately; the full value is shown only once.

Step 3: Store it as an environment variable. Alibaba's docs use DASHSCOPE_API_KEY, for example export DASHSCOPE_API_KEY="sk-..." in ~/.zshrc. Token Plan and Coding Plan subscriptions use a separate dedicated key that starts with sk-sp-; it is not interchangeable with the pay-as-you-go key.

Step 4: Point an OpenAI-compatible client at your region. The Singapore base URL is https://{WorkspaceId}.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1, and US (Virginia) uses us-east-1 in the same pattern. Alibaba asks Beijing, Singapore and Hong Kong users to migrate from the older shared domains to these workspace-specific ones.

Step 5: Protect the free quota. The quota is per model (typically 1,000,000 tokens each, input and output combined), shared by the account and its RAM users, valid 90 days and not renewed. Turn on "Free Quota Only" per model if you want calls to fail with AllocationQuota.FreeTierOnly instead of starting to bill once the quota is used.

The three open-weight rivals most often priced against Qwen each publish a USD price table. All figures are standard pay-as-you-go list prices per 1M tokens, read from each vendor's pricing page this session.

Kimi API pricing. Moonshot's platform (platform.kimi.ai/docs/pricing/chat) bills pay-as-you-go with no subscription for the API, separate from Kimi membership and Kimi Code. K3-series models have a 1,048,576-token context, cache writes priced by TTL (5 minutes or 1 hour), and batch at 60% of standard price. Moonshot states that from October 31, 2026 K2-series models are closed to newly registered users and that it plans to retire them by the end of December 2026, so new projects should price K3.

GLM API pricing. Z.ai's developer docs (docs.z.ai/guides/overview/pricing) list GLM-5.3 and GLM-5.2 at the same $1.4 / $4.4, with cached-input storage marked limited-time free. GLM-4.7-Flash and GLM-4.5-Flash are listed as free. Z.ai also sells a separate Coding Plan subscription; this page covers the per-token API only.

MiniMax API pricing. MiniMax's pricing overview (platform.minimax.io/docs/guides/pricing) shows MiniMax-M3 at a list price of $0.60 / $2.40 with a "permanent 50% off" to $0.30 / $1.20 for inputs up to 512K tokens, doubling above 512K. A Priority service tier costs 1.5x standard.

DeepSeek for reference. DeepSeek's own page lists deepseek-flash at $0.30 / $1.20 at peak on a cache miss and half that off-peak; our DeepSeek V4 Flash page covers it in depth.

Per-token prices are hard to compare until you run a real job through them. We collected 1,122 recent English X posts about these four model families in 2026 (search queries for Qwen, Kimi, GLM / Z.ai and MiniMax) and counted their tokens. The median post was 261 characters and 74 tokens; the mean was 144.6 tokens, pulled up by long posts (the 90th percentile was 300 tokens). Billing follows the mean, so we used it.

The job: summarize 10,000 posts in batches of 20, with a 60-token instruction per batch and 40 output tokens per post. That is 1.48M input tokens and 0.4M output tokens. At each vendor's list price:

What the numbers say. The spread is about 100x: $0.10 on qwen3.7-flash against $10.43 on kimi-k3. Inside Qwen alone, moving from Max to Plus cuts the job from $5.35 to $1.23. Output tokens dominate the premium models: on kimi-k3.1-pro the 0.4M output tokens cost $4.00 of the $6.95.

What it cannot show. Token counts use the cl100k_base tokenizer as a stand-in; each vendor's tokenizer splits text differently, so treat costs as estimates within a few tens of percent. Quality is not measured here, only cost.

The same sample shows how each family is discussed. Kimi posts were the most price-focused: 42.0% mentioned price, cost or tokens, and 33.7% also named a rival family or DeepSeek. GLM / Z.ai posts followed at 35.0% and 30.3%. Qwen (22.0%) and MiniMax (13.3%) posts were mostly about launches, benchmarks and usage rather than price.

Volume is uneven. The Qwen search filled 282 posts in 6.6 hours and MiniMax 240 in 8.4 hours, while the narrower GLM query needed 297.5 hours for 300 posts. Kimi's query sat between at 58.2 hours.

Official accounts. On their own timelines, @Alibaba_Qwen posted about 0.68 times a day with median 82,630 views per post, @Kimi_Moonshot 0.3 a day with median 191,584 views, @Zai_org 0.47 a day with median 151,622 views, and @MiniMax_AI 1.82 a day with median 23,870 views (last 40 to 60 posts each).

A Qwen model does not read X on its own. To summarize, classify or monitor posts you fetch them first, and in this job the data side is not small next to the model side.

With qwen3.7-flash the full job is about $1.60 on TwitterAPI.io data versus $50.10 on the official X API, so for cheap models the data source decides most of the bill. The official rate card works out to $5 per 1,000 posts against $0.15, about 33x. X's pay-per-usage tier is capped at 3M post reads a month, with Enterprise above that.

REST path. /twitter/tweet/advanced_search takes X search syntax (from:, lang:, -is:retweet, since_time:) and returns 20 posts per page with has_next_page and next_cursor. /twitter/user/last_tweets returns an account's timeline. Reads need only an API key; no X account needed. An empty search still costs the 15-credit minimum ($0.00015), so a monitor that polls every 60 seconds for a month spends $6.48 on empty checks.

MCP path. For an agent client, our hosted MCP server at https://mcp.twitterapi.io/mcp takes an Authorization: Bearer <your key> header and exposes X search and lookup as tools, so a Qwen-backed agent in an MCP-capable client can call them directly. Setup per client is on our MCP server page.

New accounts that pass the automatic bot check get up to $1 of free credits (at most 100,000 credits, no card, one per person or network, expiring after 31 days), enough for about 6,600 tweets on the search endpoint.

High-volume classification or tagging. qwen3.7-flash in the lowest tier ($0.03 / $0.13) is the lowest list price in our comparison table; at $0.10 per 10,000 posts the model cost is a rounding error next to data. Keep requests under 32K input tokens to stay in that tier.

Summaries a person will read. qwen3.7-plus ($1.23 per 10,000 posts) or qwen3.8-flash ($0.41) are the middle ground. GLM-5.3-Flash and MiniMax-M3.1-Flash land within a few cents of qwen3.8-flash.

Hard reasoning over long context. qwen3.8-max ($5.35) keeps one price up to 1M input tokens, which is simpler to budget than the tiered Plus models once prompts get long. kimi-k3.1-pro costs more per job here ($6.95) because of its $10 output price.

Many models behind one key. If you want to switch between Qwen, Kimi, GLM and DeepSeek without four accounts, a router like OpenRouter adds that convenience; compare its per-model price with the vendor's own before you commit.

Run the numbers on your own data. Our measured averages (about 145 tokens per post) apply to X posts; long-form documents and code have very different token counts. The code below logs real usage tokens per batch so you can replace our estimate with your own.

Question. What does it cost to run Qwen, Kimi, GLM and MiniMax models over real X posts, and how do developers talk about these APIs?

Accounts and queries. Four /twitter/tweet/advanced_search queries (Latest, English, no reposts) for Qwen, Kimi, GLM / Z.ai and MiniMax, up to 15 pages each; plus the timelines of the official organization accounts @Alibaba_Qwen, @Kimi_Moonshot, @Zai_org and @MiniMax_AI via /twitter/user/last_tweets, 3 pages each.

Window and sample. The most recent matching posts at collection time, collected in 2026: 1,122 search posts from 916 author slots and 220 official-account posts. 75 API requests in total.

Method. Tokens counted with the cl100k_base tokenizer as a stand-in for each vendor's own. Price and rival mentions are keyword matches on the post text. Costs use each vendor's standard list price read the same day; promotions are excluded.

What the sample cannot show. Model quality, latency or rate limits; non-English posts; tokenizer differences between vendors; and anything beyond a short collection window (the GLM query spans more days than the others).

Privacy. Aggregates only. Search authors are private individuals and are never named; only the four official organization accounts are named. Raw search pages were not stored in the repository.