Skip to content
ISSUE 001·LIVE·20:20 IL
← Journal/2026-11-16·9 min·qwen

Qwen4-Max: Alibaba's 2.4T flagship API model claims 62% lower cost per task than its predecessor

Alibaba's Qwen team released Qwen4-Max, the first model in the Qwen4 family: a 2.4-trillion-parameter sparse MoE with 71B active parameters, a 1M-token context window, thinking modes with tunable effort, and both OpenAI-compatible and Anthropic-compatible endpoints. Vendor-reported numbers show up to 49% fewer tokens on agentic coding tasks and a 62% drop in average cost per task versus Qwen3.8-Max. Input starts at 1.5 dollars per million tokens.

By Harel Asaf·AI Builder·Tel Aviv

Alibaba's Qwen team released Qwen4-Max on November 16, 2026, and it is the first model in the Qwen4 family. The short version: a 2.4-trillion-parameter sparse mixture-of-experts with 71B active parameters per token, a 1M-token context window, thinking modes with three effort levels, and access through both OpenAI-compatible and Anthropic-compatible APIs. The vendor's own numbers claim large efficiency gains over Qwen3.8-Max: up to 49% fewer tokens on Terminal-Bench 4.0 and a 62% lower average cost per task on the team's internal Work360 evaluation. API input pricing starts at $1.5 per million tokens. Everything below is sourced from the launch blog and the Alibaba Cloud and QwenCloud docs, and I label clearly what is vendor-reported.

What Qwen4-Max actually is

Qwen4-Max is built on a new architecture for the Qwen4 series. The backbone holds 2.4T parameters, with 71B activated per token, plus an additional 328B N-gram embedding parameters. Three architectural pieces matter for builders:

#### Hybrid attention for long context

Three Gated DeltaNet layers compress history into fixed-size states for every Qwen Sparse Attention layer, and a lightweight indexer selects relevant context at micro-block granularity. In practical terms, this is how the model gets to a 1M-token context window without paying full quadratic attention cost on long sequences. The team says training at this scale completed with no loss spikes, which they attribute to the architecture and optimizer changes.

#### Thinking modes with tunable effort

The model supports a thinking mode with low, high, and xhigh effort levels, plus a thinking_budget parameter that goes up to 131,072 tokens. Non-thinking output is capped at 131,072 tokens as well. Thinking tokens bill at the output rate.

#### Two API surfaces

The model is served through both an OpenAI-compatible endpoint and an Anthropic-compatible endpoint, so most existing SDK code should point at it with a base-URL change. The documented feature list includes tool calling, web search, structured output, context caching with 5-minute and 1-hour TTLs, PDF parsing, a code interpreter, and MCP support.

What it is not: the docs list no fine-tuning, no batch API, and no deployment option for this model, and the weights are not released. If your architecture depends on self-hosting or distillation, Qwen4-Max does not help you.

The benchmark claims, all vendor-reported

Every number in this section comes from the Qwen team's launch post. Treat them as marketing until third-party leaderboards catch up.

The headline comparison is against Qwen3.8-Max, the previous generation. On Work360, a benchmark the Qwen team built themselves around a taxonomy of 26 top-level domains, 182 subdomains, and 721 specialized fields, Qwen4-Max scores 52.69 versus 47.74 for Qwen3.8-Max. Token efficiency is the more interesting claim: on Work360 the model uses roughly 24% fewer tokens at xhigh effort and 35% fewer at high effort, and on Terminal-Bench 4.0 the reduction is approximately 48% and 49% respectively. Combined with lower API prices, the average cost per Work360 task drops from $4.43 to $1.67, a 62% reduction.

On secure coding, the team reports the Last Secure Code Bench rate rising from 22.0% to 31.8%, and code that is secure at all measured granularities from 8% to 14.7%. Those absolute numbers are still low, which is honest of them to publish, and worth remembering if you are evaluating models for security-sensitive code generation.

The launch post also includes a comparison table against other frontier models, with footnotes citing the Mercor APEX-Agents leaderboard, Artificial Analysis, the Arena CodeArena and AgentArena boards, the Vals Index, and Agents' Last Exam. The table itself is published as an image, so I could not verify individual figures from the text. Secondary coverage reports Qwen4-Max beating Claude Opus 5.5 on some of these tests. I would not cite specific cross-vendor numbers until the named leaderboards publish their own entries.

Pricing and access

Pricing is the part I could verify directly from the docs. Thinking and non-thinking modes share one price, with no input-length tiers:

#### Singapore (International)

Input is $1.5 per million tokens, output is $4.5 per million. Cache writes cost $1.875 per million for the 5-minute TTL and $3 per million for the 1-hour TTL, and cache reads are $0.075 per million. New users in the Singapore region get a free quota to try the model.

#### Beijing and Global

Input is $1.413 per million, output is $4.24 per million, with correspondingly lower cache prices. Global deployment is available in Frankfurt, Hong Kong, Tokyo, and US Virginia.

Two billing details worth knowing: the batch API, where supported on other models, bills at 50% of real-time pricing but is not listed as available for Qwen4-Max, and batch and cache discounts cannot be combined on the same request. The built-in web search tool also carries per-call fees on top of token costs in some regions, $10 per 1,000 calls in Singapore and roughly $0.57 in other listed regions, while the web extractor and code interpreter are free for a limited time.

Live Delivery and the operations loop

Two parts of the launch are more interesting than the benchmark table. The first is Live Delivery: for long-running tasks, the agent ships a usable first version early, then expands and improves it through successive releases while the user watches and can redirect the work at any point. The team open-sourced the Qwen-Live-Delivery-Plugin so developers can install the pattern in their own setups. This is the same continuous-delivery idea I keep coming back to with agents: intermediate artifacts you can inspect beat a black box that returns after an hour.

The second is that Alibaba Cloud says it already runs Qwen4-Max internally on operations research problems around its own infrastructure: scheduling model calls for data annotation, allocating GPU capacity online, and rescheduling CPU resources offline. The reported results are a 10.42% average reduction in annotation scheduling cost and a 4.62% increase in whole-machine idle GPU capacity duration. Vendor-reported again, but it is a concrete production loop where the model optimizes the systems that serve it.

What I would check

Run your own agentic coding workload against it through the OpenAI-compatible endpoint and measure tokens per completed task against your current model, since token efficiency is the launch's core claim. Test whether xhigh thinking effort earns its latency and cost on your tasks, or whether high is enough. If your harness is built on the Anthropic SDK, confirm the Anthropic-compatible endpoint handles your tool-calling patterns before you plan any migration. Watch the Mercor, Artificial Analysis, and Arena leaderboards over the next weeks for independent numbers. And check data residency requirements against the available regions before routing any production traffic.

What I would not conclude

Work360 is the vendor's own benchmark, so a win there is weak evidence of general superiority. The cross-vendor comparison lives in an image I could not verify line by line, and the Opus 5.5 claims should wait for the cited third-party boards. And a cheaper cost per task on one evaluation says nothing about your workload until you measure it yourself.

Frequently asked questions

What is Qwen4-Max?

Qwen4-Max is Alibaba's new flagship language model, released November 16, 2026, and the first model in the Qwen4 family. It is a 2.4-trillion-parameter sparse MoE with 71B active parameters per token, a 1M-token context window, and thinking modes with tunable effort, available only through APIs.

Is Qwen4-Max open source?

No. The weights are not released, and the docs list no self-hosting, deployment, or fine-tuning options. Access is through Alibaba Cloud Model Studio and QwenCloud. The team did open-source the Qwen-Live-Delivery-Plugin, which is a delivery-pattern plugin, not the model.

How much does Qwen4-Max cost?

In the Singapore international region, input is $1.5 per million tokens and output is $4.5 per million, with cache reads at $0.075 per million. Beijing and Global pricing is slightly lower at $1.413 input and $4.24 output. Thinking and non-thinking modes share the same price.

How big is the context window?

One million tokens of input context, with a maximum output of 131,072 tokens. The hybrid attention architecture, combining Gated DeltaNet layers with Qwen Sparse Attention, is what makes the long window practical to serve.

Can I use Qwen4-Max with my existing OpenAI or Anthropic SDK code?

Yes, in principle. The model is exposed through both OpenAI-compatible and Anthropic-compatible endpoints, so integration should be a base-URL and key change. Verify your specific tool-calling and streaming patterns before relying on it, since compatibility layers often have edge cases.

How does Qwen4-Max compare to Claude Opus 5.5?

The launch post includes a comparison table citing third-party leaderboards including Mercor APEX-Agents, Artificial Analysis, and the Arena boards, and secondary coverage reports wins over Opus 5.5 on some tests. The table is published as an image, so I could not verify the individual numbers. Treat the comparison as a vendor claim until the cited leaderboards publish their own results.

What is Live Delivery?

Live Delivery is Qwen4-Max's pattern for long-running agent tasks: the model delivers a usable first version early, then iterates through successive releases while the user can inspect progress and redirect the work at any point. Alibaba open-sourced the Qwen-Live-Delivery-Plugin so developers can adopt the pattern independently of the model.

Sources

Related reading

  • LongCat-2.5: Meituan open-sources a 1.6T agent model under MIT, trained and served without Nvidia

    Meituan's LongCat team released the weights for LongCat-2.5, a 1.6-trillion-parameter mixture-of-experts model with about 48B active parameters per token, native image understanding, a 512K-token context, and an MIT license. Training and the reference deployment run entirely on domestic Chinese accelerators, with W4A4 INT4 quantization that the team says roughly matches full-precision accuracy. Vendor benchmarks put it near Claude Opus 5 on agentic coding, below it on terminal tasks. Self-hosting needs a serious cluster.

  • Gemini 4 Hydrogen: Google's fast execution model undercuts Argon on price and beats it on intelligence

    Google launched Gemini 4 Hydrogen on November 12: a fast everyday coding and agentic model that scores 51 on the Artificial Analysis Intelligence Index at its default thinking level, above Gemini 4 Argon's 50, at 1.50 dollars per million input tokens and 7.50 per million output tokens. What the three-model Gemini 4 ladder means, how Orchestrator splits work in Antigravity, and what to verify.

  • Gemini 4 Carbon GA: what paying developers get today, and the November 30 price jump

    Google made Gemini 4 Carbon generally available on November 10: paid API access in AI Studio and Gemini for business, a /swarm multi-agent preview in Antigravity, Fairwind cybersecurity access, and introductory pricing that doubles after November 30, 2026. What changed since the October announcement, what Google claims, and what to verify yourself.

The build log

Follow the work.

New essays and prototypes, in your feed reader. Email subscriptions are paused while I connect a working delivery service. No email is collected here.

Follow via RSS