Qwen4-Max went generally available on November 16, 2026

The preview (qwen4-max-preview, Token Plan only) runs 11–25 November 2026 and is retired at the end of that window, so qwen4-max is the ID to build against. This page uses the GA figures from Alibaba Cloud's Model Studio model page (last updated November 21, 2026), its pricing page and its vision-limits page, all re-read 2026-11-21. Fable 5.2 figures come from the site registry, verified against Anthropic's pricing docs on 2026-11-11.

Claude Fable 5.1 vs Qwen 3.8 Max (previous generation)

Qwen4-Max GA · verified November 21, 2026

Claude Fable 5.2 vs Qwen4-Max

Alibaba's 2.4-trillion-parameter MoE against Anthropic's Fable-tier flagship. Qwen4-Max is roughly five times cheaper per token and takes video as input; Fable 5.2 is the better-documented model with the wider cloud footprint.

The short answer

Qwen4-Max (API model ID qwen4-max, released 2026-11-16) lists at $1.50 input / $4.50 output per million tokens in Alibaba's international region — about five times below Fable 5.2's $8/$40 — with a 1,000,000-token context, 131,072 output tokens (262,144 including chain of thought) and text, image and video input. Fable 5.2 is confirmed at 1M context, 128K output, $8/$40, with mandatory 30-day retention and availability on the Claude API, Bedrock, Google Cloud and Microsoft Foundry. Qwen4-Max is closed-weight and API-only — no self-hosting, no fine-tuning — and Alibaba publishes charts rather than a quotable benchmark table, claiming 6th place on Code Arena and a token-efficiency win on its own Work360 benchmark. Pick Qwen4-Max for price and Chinese-language work; pick Fable 5.2 for documented behaviour and enterprise data handling.

One model, several spellings

Search traffic says “Qwen 4 Max” and “Qwen Max 4”. Neither is an API model ID.

The model is Qwen4-Max and the API model ID is qwen4-max. Alibaba's Model Studio uses exactly that string in its model page, pricing table and capability matrix. The separate qwen4-max-preview ID exists only from 11 to 25 November 2026, is limited to the Token Plan, and is retired at the end of that window.

There is no bare qwen4 model ID and no qwen-max-4 model ID — the latter is a word-order flip from headlines and aggregator slugs. qwen-max-latest belongs to the older Qwen Max line and is not a Qwen4 model either; Alibaba's current model pages carry no such ID. If you need one string for a config file, it is qwen4-max.

Sources: Alibaba Cloud Model Studio model page (https://www.alibabacloud.com/help/en/model-studio/qwen4-max, last updated 2026-11-21), the qwen4-max-preview page for the preview window, and the batch-interface support list, which contains qwen4-max and no entry for qwen4 or qwen-max-4. All read 2026-11-21.

Specs and pricing side by side

Fable 5.2 from the model registry; Qwen4-Max from Alibaba Cloud's published tables. Prices are USD per million tokens in the international (Singapore) region.

SpecClaude Fable 5.2Qwen4-Max
API model IDclaude-fable-5-2qwen4-max
Released2026-11-092026-11-16
ParametersNot published2.4T total / 71B active (MoE)
Input / MTok$8.00$1.50
Output / MTok$40.00$4.50
Cache hit / MTok$0.20$0.075
Cache write, 5-minute / MTok$10.00$1.875
Cache write, 1-hour / MTok$16.00$3.00
Batch API (input / output)$4.00 / $20.00$0.707 / $2.12
Context window1M1M
Max output128K131K / 262K
ThinkingAdaptive, always on; effort defaults to highOn by default; low / high / xhigh, default high
LicenceClosed weights, API onlyClosed weights, API only — no fine-tuning, no deployment
Where it is soldClaude API, AWS, Google Cloud, Microsoft FoundrySingapore, Beijing, Frankfurt, Virginia, Tokyo, Hong Kong; Batch in Beijing only
Data retention30 days, mandatory — a Covered ModelNot confirmed

Fable 5.2 re-read from Anthropic's pricing documentation and model card on 2026-11-11; Qwen4-Max from Alibaba Cloud's model page, pricing page and vision-limits page on 2026-11-21. Qwen4-Max details that matter: the 1M context is shared between input and output, and thinking mode lowers the input ceiling to 983,616 tokens; output is capped at 131,072 tokens without thinking and 262,144 including chain of thought; the output rate already covers reasoning tokens. The 393,216-token output limit some outlets quote comes from the preview model page, not the GA page, so it is not used here. Cache writes are published at $1.875 (5 minutes) and $3 (1 hour); long-context tiers above 1M tokens are not published at all.

What the price gap means per task

List-price arithmetic on a representative coding workload — not a measured cost per task.

Worked example — 100 tasks, 10K input tokens each with a 50% cache hit, 8K output tokens each, at list rates: Fable 5.2 ≈ $0.36 per task (≈ $36.10 for 100), Qwen4-Max ≈ $0.044 per task (≈ $4.39 for 100) — about 8.2x cheaper. Qwen4-Max bills chain-of-thought tokens as output, so a reasoning-heavy trace pushes its real cost up: the ratio is a floor, not a ceiling.

Priced at international-region list rates with non-thinking output. Alibaba's Global region is slightly cheaper ($1.413/$4.24). Cache writes are $1.875 for five minutes and $3 for an hour, so long prompts amortise well on both sides, and the Batch API is $0.707/$2.12 — but on the model page re-read 2026-11-21 only Beijing marks Batch inference as supported for qwen4-max; the other five regions do not. The cost calculator has the exact math for your volumes.

Open the cost calculator

Benchmarks: what Alibaba has and has not published

One ranking, one vendor-owned benchmark, and no table of numbers.

  • Sixth place on Code Arena

    Alibaba's launch post says Qwen4-Max ranked 6th on Code Arena, a real-world coding leaderboard, and names Agents' Last Exam, Vals Index, JobBench and Chartography as further strong results. The scores themselves sit in charts, so the ranking claim is all that can be quoted.

  • Work360 is a vendor benchmark

    Alibaba describes Work360 as a proprietary benchmark covering 721 specialised fields, and says Qwen4-Max beats Qwen3.8-Max on it while using 35% fewer tokens at High effort and cutting average cost per task by around 60%. Both figures are vendor-reported on a vendor-owned benchmark with no independent replication — useful as a direction, not comparable with any third-party score.

  • What this page will not do

    It will not print a score for either model. Alibaba publishes charts rather than numbers, and Anthropic's Fable 5.2 launch table measures Fable against Opus 5.5, not against Qwen4-Max. Two vendor charts side by side would look like a comparison while measuring nothing in common.

  • Run both on your own tasks

    Both models take text and images in a chat-completions shape, both expose a high-effort reasoning mode and both accept tool definitions, so the same eval harness runs on either with a base-URL swap. Run 50–100 of your own tasks, compare cost per passing task rather than cost per token, and remember that Qwen4-Max bills thinking tokens as output.

Sources: Alibaba Cloud's Qwen4-Max press release and blog post (https://www.alibabacloud.com/blog/qwen4-max-new-ai-teammate-for-work-creativity-and-discovery_603783) and the Model Studio model page, read 2026-11-21. No third-party leaderboard, aggregated price or media figure is used on this page.

When to use which

Route on language, price, data handling and deployment reality — the four things that decide most production choices.

  • Your traffic is Chinese-language

    Qwen4-Max is Alibaba's flagship, trained and sold from a Chinese cloud with Beijing among the supported regions. For Chinese content, local compliance work or latency-sensitive mainland traffic it is the natural default. Fable 5.2 handles Chinese competently but is not positioned for mainland deployment.

  • Volume dominates the bill

    At $1.50/$4.50 against $8/$40, Qwen4-Max is roughly five times cheaper per token on published rates. At production volume that gap funds a lot of retries — but only if Qwen4-Max clears your quality bar on the same eval set, which is the part nobody can shortcut.

  • You were hoping to self-host

    You cannot. Qwen4-Max is closed-weight: Alibaba's capability matrix lists model fine-tuning and model deployment as not supported, and no Qwen4 weights are on Hugging Face. Alibaba says smaller open-weight Qwen4 models will be released, without names, dates or licence as of 2026-11-21. If open weights are a requirement, look at the Qwen3.8 line instead of this model.

  • Your input is images or video

    Qwen4-Max accepts text, image and video input and returns text. Alibaba's visual-understanding parameter table gives it 16,777,216 pixels per image, up to 2,048 images, a maximum video duration of two hours and a maximum video size of 2 GB — Qwen4-Max has its own row there, so the catch-all "other models" limits on the image-and-video page (2–40 seconds, 150 MB) do not apply to it. The two pages disagree on how many images one URL list may carry (2,048 against 80), so confirm the count in the console before relying on it. Fable 5.2 is documented as a text-and-image model.

  • You need Batch or a specific region

    qwen4-max is sold from Singapore plus Beijing, Frankfurt, Virginia, Tokyo and Hong Kong at two price tiers (International and Global). Batch support does not follow that list: on the model page re-read 2026-11-21, only the Beijing region marks Batch inference as supported; Singapore, Frankfurt, Virginia, Tokyo and Hong Kong all say not supported. Settle where your batch work runs before designing a pipeline around it.

Vendor positioning and published specs only; no independent performance measurement is claimed. Alibaba's capability matrix and price tables were re-read 2026-11-21 and can change without notice — the preview-to-GA transition moved both the output limit and the price list within five days.

Sources and verification dates

Qwen4-Max: Alibaba Cloud Model Studio model page (https://www.alibabacloud.com/help/en/model-studio/qwen4-max), model pricing page (https://www.alibabacloud.com/help/en/model-studio/model-pricing), vision limits (https://www.alibabacloud.com/help/en/model-studio/vision) and the launch post, all read 2026-11-21. Fable 5.2: Anthropic's pricing documentation (https://platform.claude.com/docs/en/about-claude/pricing), verified 2026-11-11.

Claude Fable 5.1 vs Qwen 3.8 Max

Price your own workload

A five-times price gap is real, but thinking tokens, cache writes and batch support decide what you actually pay. Run your own token mix first.

Frequently asked questions