Alibaba previewed Qwen3.8-Max — a 2.4-trillion-parameter Mixture-of-Experts flagship taking text, images, video, and documents — with promotional pricing set at 10% of the standard rate. But there is no per-token API price yet, only credit-based billing. Here is the honest breakdown.
The Preview Deal
During preview, Qwen3.8-Max runs at 10% of standard pricing through three surfaces: Alibaba’s Token Plan credit subscription, Qoder (agentic coding), and QoderWork (team workspace). An extra 80%-off discount applies to credit consumption between 22:00-08:00 (UTC+8), effectively ~0.2% of the standard rate for off-hours runs.
What “Second Only to Fable 5” Means
Alibaba positions Qwen3.8-Max as second only to Claude Fable 5. That is a claim, not a public benchmark — Alibaba published no independent benchmark table for the preview. Treat capability claims skeptically until third-party evals land.
The Catch: Credit Billing
Unlike OpenAI or Anthropic’s fixed per-token rates, Qwen3.8-Max has no published input/output pricing — you buy credits and spend them. Qwen models historically generate more output tokens per task than peers (raising real costs), and preview discounts do not survive preview end. This is the least predictable flagship to budget against.
Comparison Context
Kimi K3 ($3/$15, open weights, 2.8T) is Alibaba’s direct 2T-plus competitor and offers fixed per-token pricing. For short-term experimentation the preview is the cheapest way to touch a 2.4T model; for predictable production spend, Kimi K3 or GPT-5.6 Terra ($2/$12) are easier to forecast.
Bottom Line
Great to try this week, hard to budget for a year. Track the preview’s end date and watch for a published token rate before committing workloads.
