Alibaba's Newest AI Model Has No Per-Token Price

Models

Published 27 July 2026 | Last updated 27 July 2026

Alibaba announced Qwen3.8-Max on 19 July 2026, a 2.4 trillion parameter model it says is second only to Claude Fable 5. As of 27 July there is no published price per million tokens. The only way to reach it is a subscription starting at $6 a month, billed in Credits rather than tokens, and Alibaba's own terms prohibit using it in an application backend.

That combination is unusual for a flagship model, and it makes the cost impossible to compare against any competitor's rate card.

Does this apply to you?

If you want to try a frontier-class model inside a coding tool, this is among the cheapest routes available anywhere, and the terms permit it. Alibaba lists compatibility with Claude Code, Cursor, Qwen Code, Qoder, and OpenClaw.

If you are costing a product that calls a model from a server, this is not for you, and that is Alibaba's rule rather than an inference. There is no pay-as-you-go route to Qwen3.8-Max at all.

What the subscription actually costs

Alibaba's Token Plan Individual documentation lists three tiers, each showing a struck-through list price alongside a limited-time price.

A separate Team Edition, documented on Alibaba Cloud Model Studio, runs $30, $100, or $200 per seat per month for 25,000, 100,000, or 250,000 Credits, with a shared overage pack at $700 for 625,000 Credits.

Two mechanics matter more than the headline price. The seven-day and five-hour quotas both count every call, and reaching either one pauses the service until that window elapses. Unused quota does not carry over. And Alibaba states that the Individual plan does not currently support cancellation.

How much is a Credit?

Alibaba states that Credits consumed per request are determined by model type, token usage, thinking mode, and tool calls, and directs users to the console for actual consumption. No rate is published for Qwen3.8-Max.

One worked example exists anywhere in the documentation, and it is for a different model on the Team Edition page. Alibaba shows a single qwen3.6-plus request consuming about 3.18 Credits in total: 1.67 for 8,349 input tokens, 0.82 for 40,794 cached tokens, and 0.69 for 573 output tokens.

DERIVED Dividing Lite's 2,500 Credits by 3.18 gives roughly 790 requests across a seven-day window, or about 110 a day. That figure is an illustration of scale, not a Qwen3.8-Max rate. It uses a cheaper model, on the other edition, and Alibaba states plainly that unit prices vary by model. A 2.4 trillion parameter model with reasoning enabled would consume more.

The preview discounts make even that unstable. Alibaba states Credits consumption during the promotion is as low as 10 percent of the standard rate, with a further 80 percent off between 22:00 and 08:00 UTC+8. Both are discounts against a standard rate that has never been published.

The rule that decides who this is for

Alibaba's terms of use are explicit. Use is limited to interactive use within programming and agent tools, and using the API key for automated scripts, application backends, or non-interactive batch processing is prohibited and may result in subscription suspension or key revocation.

That single clause explains the pricing. This is not a cheap way to buy inference. It is a seat licence for a coding assistant, sold against Anthropic's and OpenAI's subscription tiers rather than against their API rates. The concurrency limits, the five-hour burst window, and the ban on backends all point the same way.

Two further conditions apply. The only available region is Singapore and the only deployment mode is Global, so prompts and outputs involve cross-border data transfer, and Alibaba states the customer is responsible for compliance. Subscriptions are personal, and sharing may result in restricted access.

Alibaba is reselling its rivals

The Individual plan's supported-model list includes Zhipu AI's glm-5.2 and DeepSeek's deepseek-v4-pro alongside Alibaba's own models. The Team Edition list goes further, adding DeepSeek v4-flash and v3.2, Moonshot's Kimi k2.7-code, k2.6 and k2.5, two more GLM versions, and MiniMax-M2.5.

Alibaba is selling access to four competitors inside its own subscription. The product being sold is the plan, not the model, which is a different business from the one a rate card describes. For context on the open-weight side of the same market, see the free models that run on local hardware.

What has not been published

Alibaba says Qwen3.8-Max is second only to Claude Fable 5. CLAIM No public benchmark table accompanies that ranking, and no independent evaluation has been published.

Also absent: a technical report, a model card, a licence, downloadable weights, and a release date for the open-weight version Alibaba says is coming. The documentation states the preview model may be taken offline or replaced with a production version when the preview ends.

The contrast with the week's other 2.4 trillion-class release is direct. Moonshot published a rate card for Kimi K3, then shipped its weights on the date it promised. Alibaba has published neither, three days earlier. Alibaba's relationship with model provenance is separately contested; Anthropic has accused it of large-scale extraction of Claude outputs, an allegation Alibaba has not accepted.

Sources

1. QwenCloud, "Token Plan Individual", retrieved 27 July 2026 (tiers, quotas, supported models, terms of use).
2. Alibaba Cloud Model Studio, "Token Plan (Team Edition) overview", last updated 1 July 2026 (seat pricing, Credits example, supported models).
3. QwenCloud, Token Plan pricing page, retrieved 27 July 2026.

Published 27 July 2026 | Last updated 27 July 2026