Models

Alibaba Cloud Ships Qwen4-Max With 25% Lower API Prices in Singapore

Alibaba Cloud released Qwen4-Max on November 16, 2026, cutting Singapore prices 25% to $1.50/$4.50 per million tokens, with a one-million-token context and shared thinking budgets.

Alibaba Cloud Releases Qwen4-Max with 25% Lower Singapore API Prices
Alibaba Cloud Releases Qwen4-Max with 25% Lower Singapore API Pricesgeralt / pixabay
By Elena Vasquez5 min read

Why it matters

  • Alibaba Cloud released Qwen4-Max on November 16, 2026, at $1.50 per million input and $4.50 per million output tokens in Singapore — 25% below Qwen3.8-Max.
  • The model has a one-million-token context window and a shared 262,144-token budget for thinking plus the final answer.
  • Cache reads cost $0.075 per million tokens; explicit cache creation costs $1.875 (five-minute) or $3 (one-hour).
  • Fine-tuning and custom deployment are unsupported in all six documented regions; batch inference runs only in Beijing.
  • Qwen4-Max-Preview, live since November 11, 2026, retires on November 25, 2026.

Alibaba Cloud cut Singapore list prices for its new flagship model by 25% on November 16, 2026, releasing Qwen4-Max through Model Studio at $1.50 per million input tokens and $4.50 per million output tokens. The predecessor, Qwen3.8-Max, lists at $2 and $6 respectively. The release puts a one-million-token context window, multimodal input and adjustable reasoning into direct price competition with the frontier-model market, where input and output rates are the main line items on every agent developer's invoice.

The price cut matters beyond Singapore's borders. Qwen4-Max is listed in Beijing, Singapore, Frankfurt, Virginia, Tokyo and Hong Kong, and Alibaba Cloud has built the release around workloads — coding agents, office productivity, frontend design and complex agentic tasks — that consume enormous token volumes. For teams running long-running agents, list price is only the beginning of the story. Cache pricing, reasoning budgets and access restrictions determine the final bill.

What does the new model actually do?

Qwen4-Max is a multimodal model. It accepts text, image and video input and produces text output, according to the model documentation. Feature support is broad: streaming, web search, parallel tool calling, function calling, prefix completion and structured JSON output are all documented.

Integration routes are familiar. Developers can call the model through OpenAI Chat Completions, OpenAI Responses, Anthropic Messages or Alibaba Cloud's own DashScope APIs. That gives existing coding and agent software several established paths to the model without vendor-specific rewrites.

How does the reasoning budget work?

The reasoning configuration is the release's most unusual mechanical detail. Thinking is enabled by default at high reasoning effort, with low, high and xhigh settings available. The default thinking budget is 131,072 tokens, matching the maximum non-thinking output.

But there's a catch that will shape real-world costs. Thinking and the final answer share one 262,144-token allowance. In other words, deliberation consumes part of the same output budget as the visible response. A model that reasons at length has less room for the answer itself, and long chains of thought eat into the output tokens a developer is paying $4.50 per million for.

What does caching cost?

The Qwen4 caching guide is where the economics get interesting. Identical input prefixes — system prompts, tool descriptions, reference documents — can be reused across requests from the same user. An explicit cache's prefix must contain at least 1,024 tokens, with five-minute or one-hour lifetimes available.

The two caching approaches behave differently:

  • Explicit caching uses breakpoints that provide deterministic hits for an identical prefix within the validity period.
  • Implicit caching identifies common prefixes automatically but does not guarantee hits.
  • Automatic caching can also place a breakpoint at the end of the input, allowing it to advance as a conversation grows.

Singapore cache-read pricing is $0.075 per million tokens for either approach — a fraction of the $1.50 standard input rate. Explicit-cache creation costs $1.875 per million tokens for a five-minute cache and $3 for a one-hour cache. Only input tokens receive the cache discount; output tokens remain at full price. For agents that repeatedly send long instructions or documents, prefix stability and cache lifetime decide whether the lower input charges actually materialize alongside the reduced list rates.

Who can use it, and how?

Access splits into two distinct routes. The first is ordinary pay-as-you-go API calls, priced as above. New Singapore users also get a one-million-token free quota, available for 90 days from the latest of Model Studio activation, model release or application approval.

The second route is subscription. The full Qwen4-Max is included in Singapore's Personal Edition Token Plan, which bills through Credits and supports interactive tools including Claude Code, Cursor, Qwen Code, Qoder and OpenClaw. The plan's terms are restrictive: it covers interactive coding and agent-tool use only. Automation scripts, custom application backends and non-interactive batch calls are prohibited. Developers building production systems that fire API calls on a schedule, or wiring the model into their own product's backend, are pushed toward pay-as-you-go.

What are the regional limits?

Regional capability varies sharply across the six listed locations:

  • Qwen4-Max is listed in Beijing, Singapore, Frankfurt, Virginia, Tokyo and Hong Kong.
  • Batch inference is supported only in Beijing.
  • Fine-tuning and custom model deployment are unsupported in all documented regions.

That last point constrains the release's practical scope. With no fine-tuning and no custom deployment, API configuration, reasoning settings and caching carry essentially all of the practical control developers have over how the managed model behaves in their applications.

What came before, and what happens next?

The full release follows a short preview window. Qwen4-Max-Preview became available at 10:00 on November 11, 2026 (UTC+8), exclusively through Token Plan. The preview is scheduled to run through November 25, 2026, after which the preview model will be retired. Preview users now have a nine-day window to migrate to the generally available model before their access route disappears.

Why the pricing structure matters more than the sticker price

Qwen4-Max's lower list rates and caching options address input costs for long-running agent workloads, but the final bill still depends on output length, reasoning depth and cache reuse. A 25% cut on list prices can evaporate if an agent reasons extensively at high effort, since thinking tokens draw from the same paid output allowance as the answer.

The access-model split compounds that. Interactive users get a cheap subscription; automation and backend workloads must run on metered API calls where every thinking token counts. With fine-tuning and custom deployment off the table in every documented region, the choices developers make about reasoning effort, prefix stability and cache lifetimes — not the headline price — will determine what Qwen4-Max actually costs them at scale.

Source: unite.ai

Share this article:

More from Elena Vasquez

Elena Vasquez

Show full bio

Market editor covering technology and product news in artificial intelligence and machine learning at AI In Context.

708 articles

Related articles

  1. Alibaba Closes Qwen4-Max Weights, Sells Flagship at $1.50 per Million Tokens
  2. Alibaba's Qwen4-Max undercuts rivals at $3.27 per benchmark task
  3. Tencent Launches Hy4 Flash at 92% Cheaper Input Tokens Than Hy4
  4. Alibaba's Qwen launches Audio 3.1 with five models, cuts prices up to 95%
  5. GPT-6.1 Terra halves API prices and closes the gap on OpenAI's biggest models

« Previous articleNext article »