Qwen4-Max vs Qwen 3.8: Price, Context, Benchmarks and What Changed

Qwen4-Max is cheaper than Qwen3.8-Max ($1.50/$4.50 vs $2/$6 per 1M tokens), keeps the 1M context, triples the output limit and scores higher on Alibaba's benchmarks. Here's the full comparison and how to test both yourself.

INEZA Felin-Michel

INEZA Felin-Michel

16 November 2026

Qwen4-Max vs Qwen 3.8: Price, Context, Benchmarks and What Changed

Apidog for Enterprise

On-Premises Deploy

SSO & RBAC

SOC 2 Compliant

Explore Apidog Enterprise

Qwen4-Max vs Qwen 3.8 comes down to this: Qwen4-Max costs less and does more. It’s $1.50 input and $4.50 output per million tokens against Qwen3.8-Max’s $2 and $6, with the same 1M-token context, a 393K output limit against 131K, and cached input at less than a third of the price. Alibaba’s own benchmarks put it ahead on professional work while using fewer tokens.

That’s the picture on November 16, 2026, after Qwen4-Max moved from preview to general release. Below: the specs side by side, what changed, the price math, what’s still vendor-reported, and a hands-on test you can run on both. New to either model? Start with what Qwen4-Max is and what Qwen 3.8 is.

Qwen4-Max vs Qwen 3.8 at a glance

Every row comes from Alibaba’s own pages, checked November 16, 2026.

Qwen4-Max Qwen3.8-Max
Status Generally available, November 2026 Generally available since August 2026
Model ID qwen4-max qwen3.8-max
Input price $1.50 per 1M tokens $2 per 1M tokens
Output price $4.50 per 1M tokens $6 per 1M tokens
Cached input $0.075 per 1M tokens $0.25 per 1M tokens
Context window 1M tokens 1M tokens
Max output 393K tokens 131K tokens
Inputs Text, images and video Text, images and video
Rate limit 5,000,000 tokens per minute 2,000,000 tokens per minute
Parameters 2.4T backbone, 71B activated per token 2.4T total, Mixture-of-Experts
API protocols OpenAI- and Anthropic-compatible OpenAI- and Anthropic-compatible
On Token Plan Yes Yes, “40% Off Nights”

Sources: the Qwen4-Max model page, the Qwen3.8-Max model page and the Token Plan pricing page.

Comparison diagram of Qwen4-Max and Qwen3.8-Max: input price $1.50 vs $2, output price $4.50 vs $6, cached input $0.075 vs $0.25, max output 393K vs 131K, and the same 1M-token context window

What changed in Qwen4-Max

The launch post describes a new architecture and reports gains across the board. These are the Qwen team’s own results.

1. A new architecture at the same scale. Both models sit at 2.4 trillion parameters. Qwen4-Max rebuilds the sparse Mixture-of-Experts design with Gated DeltaNet layers and Qwen Sparse Attention for long sequences, a Gated Residual design for stable training, and N-gram Embedding for extra capacity. Alibaba says its base model activates 25% fewer parameters than Qwen3.8-Max’s and scores 5% to 13% better on its in-house knowledge, code, QA and STEM evaluations.

2. Higher scores on professional work. On Work360, Alibaba’s benchmark of industry tasks across 26 domains, Qwen4-Max scores 52.69 against 47.74 for Qwen3.8-Max.

3. Fewer tokens per task. On Work360, token use drops about 24% at Xhigh effort and 35% at High effort. On Terminal-Bench 4.0, it drops about 48% to 49%.

4. Safer code. On Last Secure Code Bench, the function-level pass rate for working, secure code rose from 22.0% to 31.8%. The share of code generated securely at function, file and repository level rose from 8% to 14.7%.

5. Longer runs. Qwen 3.8’s model page already described autonomous projects “spanning 10+ days.” Qwen4-Max adds Live Delivery, where the agent ships a usable first version early and improves it release by release while you steer. The team open-sourced it as the Qwen-Live-Delivery-Plugin.

The direction hasn’t changed between generations. Qwen4-Max goes further down the same road: coding, professional work and long agentic tasks.

Price: Qwen4-Max vs Qwen 3.8

Qwen4-Max is 25% cheaper on both input and output, and 70% cheaper on cached input.

Per 1M tokens Qwen4-Max Qwen3.8-Max
Input $1.50 $2.00
Output $4.50 $6.00
Implicit cache input $0.075 $0.25

The cache rate matters most for agents. A coding agent resends the conversation, tool schemas and file contents on every turn, so most of its input is cache hits. Take a session with 10 million cached input tokens, 1 million fresh input tokens and 500,000 output tokens:

That’s 40% less for the same tokens. Then the token savings stack on top: if Qwen4-Max needs fewer tokens for the job, as Alibaba reports, the gap widens. Alibaba’s own figure is a drop in average cost per Work360 task from $4.43 to $1.67, a 62% cut.

Both models are also on the Token Plan, Alibaba’s subscription for coding tools, where they draw from the same Credit pool. See our Qwen4-Max Token Plan guide for tiers and rules, and the Qwen 3.8 API guide for the older model’s full price sheet.

What’s still vendor-reported

The preview left most comparison cells empty. The general release filled them, with three caveats.

For access details on every route, see how to use Qwen4-Max.

Run your own Qwen4-Max vs Qwen 3.8 test

Vendor benchmarks point in a direction. Your own test settles it. API work fits well, because an endpoint either passes its tests or it doesn’t.

Both models run on the same endpoints, so switching is one model name. With Qwen Code on the Token Plan, list both in ~/.qwen/settings.json:

{
  "$version": 4,
  "env": { "QWENCLOUD_TOKEN_PLAN_API_KEY": "YOUR_API_KEY" },
  "modelProviders": {
    "openai": [
      { "id": "qwen4-max", "baseUrl": "https://token-plan.maas.qwencloudapi.com/compatible-mode/v1", "envKey": "QWENCLOUD_TOKEN_PLAN_API_KEY" },
      { "id": "qwen3.8-max", "baseUrl": "https://token-plan.maas.qwencloudapi.com/compatible-mode/v1", "envKey": "QWENCLOUD_TOKEN_PLAN_API_KEY" }
    ]
  }
}

Then set up the fair test:

  1. Write the contract once. Create one openapi.json for a small API, for example POST /orders and GET /orders/{id} with validation and auth rules. Neither model writes the spec, so neither model grades its own homework.
  2. Make two empty folders, build-qwen4 and build-qwen38, each with a copy of the spec.
  3. Send the identical build prompt in each folder, with qwen4-max selected in one and qwen3.8-max in the other: “Build the API in openapi.json with Express. Run it on port 3001 (or 3002). Don’t change the spec.”
  4. Record the basics: time to a running server, tokens used, and whether you had to step in.

Then test both builds against the same suite. That’s where Apidog CLI comes in.

Score both builds with Apidog CLI

Your coding agent drives Apidog CLI: it imports the shared spec, creates one set of test cases from it, and runs that set against each build. Same tests, two targets, two reports.

Workflow diagram: one OpenAPI spec feeds two builds, one written by Qwen4-Max on port 3001 and one by qwen3.8-max on port 3002; Apidog CLI runs the same test cases against both environments and produces two pass or fail reports side by side

Set up Apidog CLI once

npm install -g apidog-cli@latest
apidog auth login --with-token <YOUR_APIDOG_TOKEN>
apidog skill install

apidog skill install teaches your agent the CLI. The installation guide covers tokens.

Give your agent one prompt

Any agent can referee; it doesn’t have to be a Qwen model.

Import ./openapi.json into Apidog project <projectId>. Create two environments:
qwen4 at http://localhost:3001 and qwen38 at http://localhost:3002. From the spec,
create test cases for POST /orders: 201 on a valid order, 400 on a missing
customerId, 401 without a token, plus a GET /orders/{id} 404 case. Run every case
against both environments, save JSON reports, and give me a pass/fail table per model.
Don't modify either build.

The apidog commands it runs

Every flag below comes from apidog <command> --help on Apidog CLI 2.2.12:

apidog import --project <projectId> --format openapi --file ./openapi.json
apidog environment create qwen4 --project <projectId> --base-url http://localhost:3001
apidog environment create qwen38 --project <projectId> --base-url http://localhost:3002
apidog cli-schema get test-case-create
apidog cli-schema validate test-case-create --file ./case-201.json
apidog test-case create --project <projectId> --file ./case-201.json
apidog run --test-case <caseId> --project <projectId> -e <qwen4EnvId> -r cli,json --out-file qwen4-{GENERATE_TIME}
apidog run --test-case <caseId> --project <projectId> -e <qwen38EnvId> -r cli,json --out-file qwen38-{GENERATE_TIME}

cli-schema get hands the agent the exact JSON shape for a test case, and cli-schema validate checks each file before anything is written. The agent repeats test-case create and the two apidog run lines for each case. The cli reporter prints a pass/fail summary, and json saves a report in apidog-reports, which the agent reads to build its table. Add --upload-report to view a run in the Apidog app.

What to read in the reports

Count three things per model: cases passed on the first build, fix rounds needed to reach all green, and any case where the model changed the contract instead of the code. Run the whole test two or three times, since one run tells you little. For more agent-driven patterns, see the Apidog CLI agent workflow and Apidog CLI in Claude Code. Apidog’s Free plan covers up to 4 users.

Which should you use?

Match the model to the job:

Keep the model ID in one config value, so the next swap is a one-line change. See also Qwen 3.8 for coding and our guide to the best Qwen models.

FAQ

Is Qwen4-Max better than Qwen 3.8? On Alibaba’s numbers, yes: 52.69 against 47.74 on Work360, with fewer tokens per task and a higher secure-coding pass rate. Those are vendor results, so test both on your own task.

Can I call Qwen4-Max through the API like Qwen 3.8? Yes. Both are on QwenCloud’s pay-per-token API with OpenAI- and Anthropic-compatible endpoints. The model IDs are qwen4-max and qwen3.8-max.

Does Qwen4-Max have a 1M-token context like Qwen 3.8? Yes. Both list a 1M context. Qwen4-Max allows up to 393K output tokens; Qwen3.8-Max allows 131K.

How much does Qwen4-Max cost compared with Qwen 3.8? Qwen4-Max costs $1.50 input and $4.50 output per million tokens. Qwen 3.8 costs $2 and $6. Cached input is $0.075 against $0.25.

Should I switch my app from Qwen 3.8 to Qwen4-Max? Probably, after a test. It’s cheaper and scores higher on Alibaba’s benchmarks, and the API is compatible. Run your evaluation suite on both before you change production.

Explore more

What Is Firecrawl Local? Run Firecrawl on Your Own Machine

What Is Firecrawl Local? Run Firecrawl on Your Own Machine

Firecrawl Local is one Rust binary that scrapes, crawls and maps websites to markdown on your own machine and serves the Firecrawl v2 API on localhost:3002. Here's what runs locally, what goes to the cloud, how to install it and where its limits are.

19 November 2026

What Is Gemini 4 Neon?

What Is Gemini 4 Neon?

Gemini 4 Neon is Google's middle Gemini 4 model, released Nov 18, 2026, for complex coding and analysis that runs often: $3/$15 per 1M tokens, a shared 1M-token context, four thinking levels and no free API tier.

19 November 2026

What Is GPT-6.2 Astra?

What Is GPT-6.2 Astra?

GPT-6.2 Astra is OpenAI's new frontier model for hard, long tasks; GPT-6.2 Sol is the everyday model at a quarter of the price. Here is what each one is, what it costs, where you can use it and how to choose.

18 November 2026

Practice API Design-first in Apidog

Discover an easier way to build and use APIs

Qwen4-Max vs Qwen 3.8: Price, Context, Benchmarks and What Changed