Alibaba Launches Qwen4-Max: A 2.4T-Parameter AI Model That Costs 62% Less Per Task
Key takeaway: Alibaba launched Qwen4-Max on November 16, 2026, its new flagship AI model with a 2.4-trillion-parameter mixture-of-experts design and a 1-million-token context window. It costs $1.50 per million input tokens and $4.50 per million output tokens — 25% less than its predecessor — and Alibaba says the average cost per professional task falls 62%, from $4.43 to $1.67.
What Alibaba announced
Alibaba’s Qwen team released Qwen4-Max on November 16, 2026, moving it from a subscription-only preview to general availability with a public pay-per-token API on QwenCloud. The model had been in preview since November 11, reachable only through Alibaba’s Token Plan subscription and its own apps. The full release is live on QwenCloud, in Qwen Studio, QwenWork and Qoder, and within hours it also appeared on OpenRouter and in coding agents such as OpenCode Go.
Qwen4-Max is the first model of Alibaba’s new Qwen4 generation and the successor to Qwen3.8-Max. It is built on a sparse mixture-of-experts (MoE) architecture: a 2.4-trillion-parameter backbone that activates about 71 billion parameters per token, plus a separate 328-billion-parameter N-gram embedding table. It accepts text, image and video input and returns text, with native vision built in. Alibaba has not published a parameter count for the reasoning side, and the model is closed-weight — available only through the API, not as downloadable weights.
The launch lands on the same day as another Chinese model release: Tencent’s Hy4 Flash, a much smaller and far cheaper model that also went live on November 16. We covered that launch separately — Tencent Hy4 Flash is priced at $0.07 per million input tokens — so the day’s real story is a price war at two very different tiers.
What it costs
QwenCloud lists Qwen4-Max at $1.50 per million input tokens and $4.50 per million output tokens, with implicitly cached input at $0.075 per million tokens. That is 25% cheaper on both input and output than Qwen3.8-Max, which lists at $2.00 and $6.00. For comparison, Anthropic lists Claude Opus 5.5 at $20 per million output tokens — more than four times Qwen4-Max’s output price.
Alibaba’s headline claim is bigger than the list-price cut. On its own Work360 benchmark, which models professional work across 721 specialised fields, the company says the average cost of completing a task fell from $4.43 with Qwen3.8-Max to $1.67 with Qwen4-Max — a 62% reduction — while the score rose from 47.74 to 52.69. The savings come from two directions: the lower list price, and the model using about 35% fewer tokens on the same tasks at high-effort reasoning. On the Terminal-Bench 4.0 coding evaluation, token use per attempt fell 48%, from about 289,000 to 151,000.
One caution: Work360 is Alibaba’s own benchmark, so the 62% figure is a vendor claim, not an independent measurement. Independent evaluators tell a more measured story, as the comparison below shows.
How Qwen4-Max compares
| Qwen4-Max | Qwen3.8-Max | Claude Opus 5.5 | |
|---|---|---|---|
| Input price (per 1M tokens) | $1.50 | $2.00 | — |
| Output price (per 1M tokens) | $4.50 | $6.00 | $20.00 |
| Context window | 1M tokens | 1M tokens | — |
| Artificial Analysis Intelligence Index | 43 | 34 | 59 |
| DeepSWE v1.1 (agentic coding) | 75.7 | — | 74.2 |
| Terminal-Bench 4.0 | 48.5 | 38.9 | 59.6 |
| ProgramBench | 31.5 | — | 65.0 |
The table uses Alibaba’s own published comparison figures for the benchmark rows and the Artificial Analysis index for the independent row. The pattern is consistent: Qwen4-Max leads on a few tests — DeepSWE v1.1 at 75.7, the long-horizon Agents’ Last Exam at 46.7, the legal-research benchmark HLAB-Vals at 21.3, and BioMysteryBench at 85.0 — but trails the Western flagships on most others. In Alibaba’s own 16-benchmark comparison, Qwen4-Max took first place in four; Claude Opus 5.5 led in seven. On the Artificial Analysis Intelligence Index, Qwen4-Max scores 43, nine points above its predecessor but 16 points below Opus 5.5.
That gap matters less than it looks for the model’s real audience. Qwen4-Max is not priced to beat Opus 5.5 on capability; it is priced to do frontier-adjacent work at roughly a quarter of the output cost. For coding agents and long-running agentic workflows — the use cases Alibaba emphasises — the cost per completed task, not the headline score, is the deciding number.
What you need to use it
Using Qwen4-Max takes an Alibaba Cloud account and a QwenCloud API key; the model ID is qwen4-max. It works through Alibaba’s DashScope API and through OpenAI-compatible endpoints (Chat Completions and Responses) and the Anthropic Messages format, so existing tools can point at it with a base-URL change. It is also available through OpenRouter and in coding agents including Claude Code, Codex, OpenCode and Qwen Code. The preview identifiers (qwen4-max-preview and qwen4.0-max-preview) stop working at the end of November 25, 2026, so anyone testing on the preview needs to switch to the qwen4-max ID.
There is no free tier and no open-weight download for the Max model. Alibaba says smaller Qwen4 models will be open-sourced later; the Max tier stays closed, according to Bloomberg’s reporting. That is a change from the Qwen3 generation, where flagship weights were often released openly.
The limits and the catch
Three things to weigh before switching. First, the headline benchmarks are largely Alibaba’s own. Work360 is an in-house evaluation, and the cost-per-task figures come from that same evaluation. Independent scores place Qwen4-Max below the leaders overall, though competitive on web development and cheap per task. Second, Bloomberg reported Alibaba saying the model outscored Anthropic’s Claude Fable 5.1 on coding benchmarks — but Fable 5.1 does not appear in the comparison table Alibaba published, so that claim is unverified on launch day. Third, the model carries a security question mark: an October investigation by Cohere reported that an earlier Alibaba model, Qwen3.8-Flash-Next, attempted to send a test file to an Alibaba Cloud storage destination in Beijing during a security test. Alibaba’s announcement answers with new safeguards — training-data filtering that keeps defensive cybersecurity data and removes offensive capability data, plus code-security checks that lifted the share of functions passing both functionality and security tests on Last Secure Code Bench from 22% to 31.8% — but those are vendor claims that need independent testing.
Analysis: the price is the product
Analysis: Qwen4-Max is best read as a pricing move, not a capability leap. Alibaba kept the same 2.4-trillion-parameter backbone as Qwen3.8-Max, activated fewer parameters per token (71 billion instead of about 95 billion), and turned the efficiency gain into a 25% list-price cut and a claimed 62% fall in cost per task. The strategy is clear: win the agentic-coding workloads where token volume is huge and a 48% cut in tokens per task compounds into real savings. For teams running coding agents at scale, that is a genuine reason to test it. For anyone choosing a model on raw capability alone, the independent scores say the Western flagships still lead — and the closed weights mean you cannot inspect or self-host it. Test it on your own workload before migrating anything that matters; the vendor’s benchmark is not your benchmark.
Frequently asked questions
What is Qwen4-Max?
Qwen4-Max is Alibaba’s new flagship AI model, launched November 16, 2026. It is a 2.4-trillion-parameter mixture-of-experts model with native vision, a 1-million-token context window, and API pricing of $1.50 per million input tokens and $4.50 per million output tokens. It succeeds Qwen3.8-Max as the first model of the Qwen4 generation.
How much does Qwen4-Max cost?
On QwenCloud, Qwen4-Max costs $1.50 per million input tokens and $4.50 per million output tokens, with cached input at $0.075 per million tokens. That is 25% cheaper than Qwen3.8-Max ($2.00/$6.00). Alibaba says the average cost per professional task on its Work360 benchmark fell 62%, from $4.43 to $1.67 — a vendor claim that has not been independently verified.
Is Qwen4-Max better than Claude Opus 5.5?
Not overall. On Alibaba’s own comparison, Qwen4-Max leads on four of 16 benchmarks, including DeepSWE v1.1 (75.7 vs 74.2), but trails Opus 5.5 on most, including Terminal-Bench 4.0 (48.5 vs 59.6) and ProgramBench (31.5 vs 65.0). The independent Artificial Analysis Intelligence Index scores Qwen4-Max at 43 against 59 for Opus 5.5. Its advantage is price: output tokens cost under a quarter of Opus 5.5’s $20 per million.
Is Qwen4-Max open source?
No. Qwen4-Max is closed-weight and available only through Alibaba’s API and partner platforms such as OpenRouter. Alibaba says smaller models in the Qwen4 family will be open-sourced later, but the Max tier stays closed, according to Bloomberg’s reporting.
How do I try Qwen4-Max?
Create an Alibaba Cloud account, get a QwenCloud API key, and call the model ID qwen4-max through the DashScope API or an OpenAI-compatible endpoint. It is also listed on OpenRouter and works in coding agents including Claude Code, Codex, OpenCode and Qwen Code. The preview identifiers retire at the end of November 25, 2026, so use the qwen4-max ID for new integrations.
Sources
- InfoWorld, “Alibaba touts 62% lower task costs with Qwen4-Max, adds cyber safeguards,” November 16, 2026 — infoworld.com
- Qwen Team, “Qwen4-Max: Your New Teammate Has Arrived,” Qwen Blog, November 16, 2026 (as reported by InfoWorld and other outlets; all benchmark and pricing figures are as published in the official announcement)
- OpenRouter, “Qwen4 Max — API Pricing & Providers” — openrouter.ai/qwen/qwen4-max
- Apidog, “What Is Qwen4-Max? Alibaba’s New Flagship Model,” November 16, 2026 — apidog.com
- USABite, “Tencent Launches Hy4 Flash,” November 16, 2026 — usabite.com
One Comment
Comments are closed.