Qwen4-Max API Pricing vs the Coding-Bench Gap

Night desk with keyboards, a locked metal box, and a blurred whiteboard.

Qwen4-Max API pricing is $1.50 per million input tokens and $4.50 per million output on QwenCloud’s Singapore region, for a hosted-only model Alibaba released on November 16, 2026. That rate undercuts GPT-6.1 Sol’s $2 and $10 list prices and Terra’s $5 output, but not Terra’s $1 input. Closed weights mean the client swap is not an open fallback, and hard coding scores still trail Claude Opus 5.5.

Updated November 16, 2026

Key Takeaways

  • Qwen4-Max is hosted only: 2.4T MoE, 71B active, 1M context.
  • Singapore list price is $1.50 input and $4.50 output per million.
  • MarkTechPost and WebProNews put Work360 cost at $4.43 to $1.67.
  • ProgramBench is 31.5 versus 65.0 for Claude Opus 5.5.

What does Qwen4-Max API pricing actually buy?

On QwenCloud’s Singapore region, WebProNews reports $1.50 per million input tokens and $4.50 per million output, with cache reads at $0.075. Beijing and global deployments run slightly lower, at $1.413 and $4.24, and thinking and non-thinking modes cost the same. MarkTechPost says those rates fell from $2 and $6 on Qwen3.8-Max, with implicit cache reads down 70% from $0.25.

Tokens move the bill as much as the rate card. On Work360, Alibaba’s benchmark across 721 specialized fields, use fell about 35% at High effort and about 24% at Xhigh. Alibaba Cloud’s announcement said average cost per task fell by around 60%. MarkTechPost and WebProNews put that drop at $4.43 to $1.67; MarkTechPost calculates 62%.

On Terminal-Bench 4.0, token use fell about 48% to 49% even where the score lags.

Model Input / 1M Output / 1M ProgramBench Terminal-Bench 4.0
Qwen4-Max (Singapore) $1.50 $4.50 31.5 48.5
Claude Opus 5.5 Not reported Not reported 65.0 59.6
Kimi K3.1 Not reported Not reported Not reported 67.2
GPT-6.1 Sol $2 $10 Not reported Not reported
GPT-6.1 Terra $1 $5 Not reported Not reported

Opus and Kimi list prices are not in these sources. Bench cells follow Alibaba’s chart as WebProNews reported it. GPT-6.1 Sol and Terra are OpenAI list rates from that same report, input then output.

How wide is the coding-bench gap?

MarkTechPost puts Qwen4-Max first on 4 of 16 benchmarks, including DeepSWE v1.1 at 75.7 versus 75.5 for Kimi K3.1. Agents’ Last Exam v1.1 is 46.7 versus 43.4 for Opus 5.5 and Kimi K3.1. WebProNews puts DeepSWE just ahead of Opus 5.5 at 74.2. HLAB-Vals-1.1 is 21.3 versus 11.7 for Kimi K3.1 and 9.2 for Opus 5.5. MarkTechPost says Work360 rose from 47.74 to 52.69.

  • ProgramBench: 31.5 versus 65.0 for Opus 5.5.
  • Terminal-Bench 4.0: WebProNews reports 48.5 versus 59.6 for Opus 5.5 and 67.2 for Kimi K3.1 on Alibaba’s chart.
  • Code Arena: Alibaba said 6th; WebProNews reports No. 6 on the WebDev board at a preliminary 1,750.

MarkTechPost notes those chart scores were not independently reproduced at launch. It cites Moonshot’s card at 65.7 for Kimi K3.1 Pro and 66.4 for Opus 5.5 on Terminal-Bench 4.0, and states Kimi at 67.2 on Alibaba’s chart. It does not cite 59.6.

The ProgramBench spread means a cheaper token is not a cheaper finished coding task. Misses still add review, the load already tracked in coding-agent task completion.

Outside that chart, Arena ranked the model No. 14 on Agent Arena and the top Chinese model, with Alibaba fifth among labs, behind Anthropic, Google, OpenAI, and Meta. Vals AI ranked it 11th on the Vals Index, also the highest Chinese model, with coding and finance gains over Qwen3.8-Max.

What does an Anthropic-compatible client change?

The model ID is qwen4-max. QwenCloud works with OpenAI-compatible and Anthropic-compatible clients, including function calling, structured outputs, context caching, and web search, at a default 5,000,000 tokens per minute. Thinking mode is toggled with enable_thinking. Access also includes the Token Plan, Qwen Studio, QwenWork, and Qoder.

WebProNews says it is live on OpenRouter and Vercel’s AI Gateway, served from Singapore, Beijing, Frankfurt, Hong Kong, Tokyo, and Virginia. For Claude-shaped SDKs, that is a client swap, not a published claim of Claude Code parity.

The write-ups already disagree. MarkTechPost lists fine-tuning, batch processing, and a 393K max output. WebProNews, citing developer docs, says fine-tuning and batch inference are unsupported and caps thinking plus the answer at 262,144 tokens. Confirm the billed endpoint before treating compatibility as a drop-in.

Specs that do agree: a 1 million-token window, native image and video input, and 71 billion active parameters per token — 25% fewer than Qwen3.8-Max at the same 2.4T scale. MarkTechPost also reports 328 billion N-gram embedding parameters on top of that backbone.

Why does closed-weight lock-in still matter?

No weights means no self-hosting and no quantized local build of this flagship. That reverses the August open release of Qwen3.8-2.4T-A95B, a text-only model and the first Max-class Qwen with open weights. Alibaba says smaller Qwen4 models will ship open-weight soon. WebProNews says no date has been given.

An Anthropic-compatible rate cut still leaves this checkpoint on Alibaba’s side of the API. Teams comparing coding-agent budgets should price the bench gap into cost per finished task, not only per million tokens.

Frequently asked questions

Is Qwen4-Max available with open weights?

No. It is hosted only, with no self-hosting or quantized local build. Alibaba said smaller Qwen4 models will get open weights later, and WebProNews reports no date.

What is Qwen4-Max API pricing in Singapore?

WebProNews reports $1.50 per million input tokens and $4.50 per million output, with cache reads at $0.075. Alibaba said Work360 average cost per task fell by around 60%; MarkTechPost and WebProNews put that at $4.43 to $1.67.

Does the Qwen4-Max API work with Anthropic clients?

MarkTechPost says the QwenCloud API works with Anthropic-compatible and OpenAI-compatible clients. Published write-ups still disagree on fine-tuning, batch inference, and maximum output.

How far behind is Qwen4-Max on coding benchmarks?

WebProNews, citing Alibaba’s chart, reports 31.5 on ProgramBench versus 65.0 for Claude Opus 5.5, and 48.5 on Terminal-Bench 4.0 versus 59.6 and 67.2. MarkTechPost says it led DeepSWE v1.1 at 75.7.

Notice an error?

Help us improve our content by reporting any issues you find.