Key takeaways
- Qwen announced Qwen4-Max on November 16, 2026, with a one-million-token context window and native vision; QwenCloud API and Token Plan access are announced now.
- OpenRouter lists qwen/qwen4-max with text, image and video input, text output, tool calling, structured outputs and reasoning enabled by default.
- Listed OpenRouter rates are $1.50 per million input tokens and $4.50 per million output tokens. These are registry prices, not a verified invoice or a universal QwenCloud tariff.
- Qwen’s efficiency and productivity results are vendor-reported. Smaller Qwen4 models are coming to open source; that promise does not make Max an open-weight release.
What launched on November 16
Alibaba’s Qwen team announced Qwen4-Max on November 16, 2026, introducing a new flagship with native vision and a one-million-token context window. Qwen says the model is available through QwenCloud’s API and Token Plan, and can also be used in QwenWork, Qoder and QwenStudio. OpenRouter separately lists the API identifier qwen/qwen4-max at $1.50 per million input tokens and $4.50 per million output tokens.
The immediate news for builders is another large-context multimodal model with configurable reasoning and agent-oriented interfaces—not downloadable weights. Qwen says smaller models in the family are coming to open source. The announcement does not establish a Max weight release, a Max download license or self-hosting requirements.
Access needs a qualification: during our review, OpenRouter’s model page showed one provider, Alibaba Cloud Int., with a “Down” indicator. Its public endpoints API separately returned an endpoint with reported uptime values of 100. Those inconsistent snapshots do not establish dependable availability. We verified the listing, not a successful inference request; check current provider status before starting a trial.
What the model and API expose
Qwen describes a sparse mixture-of-experts architecture with a 2.4-trillion-parameter backbone, 71 billion parameters activated per token and a further 328 billion N-gram Embedding parameters. The team attributes efficiency improvements to Gated DeltaNet and Qwen Sparse Attention layers, gated residual connections and context-based embedding lookups. These are the developer’s architecture disclosures, not independently inspected model weights.
OpenRouter’s registry describes text, image and video inputs with text output. That is not native video generation or audio output. It lists a 1,000,000-token context and a top-provider maximum completion of 393,216 tokens; the completion ceiling is not a promise that every request can use both maxima simultaneously.
The listing supports tools, tool choice, JSON-style response formats and structured outputs. Reasoning defaults to high, with low and xhigh options and a way to disable reasoning. These are supported interface settings, not evidence that the model always chooses the right tool, follows every schema or completes a long-running task correctly.
Pricing and the efficiency claim
The retrieved OpenRouter registry quotes $0.075 per million cache-read tokens and $1.875 per million cache-write tokens alongside the $1.50 input and $4.50 output rates. These are that route’s listed prices as retrieved on November 16, not a verified QwenCloud regional price schedule. Cache eligibility, account terms, reasoning usage and actual response length can affect the bill.
Qwen says its Work360 evaluation’s average task cost fell from $4.43 with Qwen3.8-Max to $1.67 with Qwen4-Max, a vendor-reported 62% reduction. It also reports lower token usage at high and xhigh effort on Work360 and Terminal-Bench 4.0. This combines the vendor’s workload, model behavior and prices; it is not evidence that an existing customer’s total bill will fall by the same percentage.
For an initial comparison, keep the task, tool harness, reasoning setting and acceptance criteria fixed. Record completed-task cost, failed attempts and elapsed time rather than comparing token prices alone. A million-token window can be useful, but supplying more context is neither free nor proof of reliable recall across that window.
Why the agent claims matter—and what is unverified
The launch emphasizes professional tool use, long-horizon execution and finished deliverables. Qwen reports a Work360 score of 52.69 versus 47.74 for Qwen3.8-Max, and describes the benchmark as covering a taxonomy of 26 top-level domains. These are vendor-reported results. AccessAllGPT did not rerun the suite or independently validate the reported task outcomes.
Qwen’s Live Delivery approach is meant to produce a usable first version early, then refine it through successive releases with user guidance. The team also announces an open-source Qwen-Live-Delivery-Plugin. The model launch and an agent product are different layers: reproducing a vendor demonstration can depend on its harness, tools, assets and permissions, not just the model ID.
Our analysis: the practical reason to evaluate Max is the combination of multimodal context and iterative tool-assisted work. It is not a reason to hand a new model unrestricted production access. Start with a bounded, non-sensitive document or design task whose outputs can be checked; keep consequential actions subject to approval.
What to do next
If you already buy Qwen APIs, confirm which QwenCloud endpoint and plan expose Max in your account, then compare it with your current model on a small set of representative tasks. OpenRouter users can start with qwen/qwen4-max once the provider is healthy, explicitly selecting reasoning effort so the comparison does not silently change compute budget.
If you require self-hosting, wait for actual released artifacts and their licenses rather than treating the promise of smaller models as deployable software. No hands-on test was performed for this paid model, and this briefing makes no latency, accuracy or production-readiness recommendation.
A launch-day evaluation card
A proposed trial checklist, not an AccessAllGPT benchmark result.
Entries stay in this browser tab and are not submitted to AccessAllGPT. Blank responses are copied as [Unresolved].
Verify a successful non-sensitive request on your intended provider; the model page and endpoint-status snapshots differed during review.
Test your actual image/video inputs, tool schemas and long-context questions. Do not infer generation modalities from input support.
Hold reasoning effort and tool harness fixed; compare cost per accepted result, including retries and cache behavior.
Use managed access for evaluation. Wait for real artifacts and licenses if open weights are a requirement.
Primary sources
Browse the publication-wide evidence index →
- Qwen4-Max: Your New Teammate Has ArrivedQwen Team · Reviewed: November 16 publication date; introduction; token efficiency; pre-training; Work360; Live Delivery; model-harness-product co-design; how to use; summary · Retrieved · Supports: First-party launch date, architecture disclosures, million-token context, vendor-reported efficiency and benchmark results, announced QwenCloud/product access and future smaller open-source models. Biological and wet-lab applications are excluded from this briefing.
- Qwen4 Max — pricing and providersOpenRouter · Reviewed: Model description; November 16 release label; input/output pricing; modalities; provider row and status indicator · Retrieved · Supports: Registry route, listed prices, input/output modalities, reasoning controls and observed provider-page availability caveat; not proof of a successful call.
- OpenRouter public model registryOpenRouter · Reviewed: qwen/qwen4-max entry: created timestamp, architecture, context_length, pricing, top_provider, supported_parameters and reasoning · Retrieved · Supports: Machine-readable identifier, million-token context, per-token price values, completion ceiling, tool/structured-output parameters and reasoning defaults.
- Qwen4 Max public endpointsOpenRouter · Reviewed: Alibaba endpoint entry, pricing, max_completion_tokens, supported parameters and uptime fields · Retrieved · Supports: Endpoint listing and reported status/uptime snapshot; its disagreement with the model-page Down indicator is disclosed rather than treated as verified availability.
Limitations
Source-backed launch briefing, not hands-on testing. No paid API request, independent benchmark, weight inspection or deployment was performed. Architecture, productivity and cost comparisons are vendor claims. OpenRouter status surfaces disagreed during review; no stable call availability is established. Regional access, QwenCloud account-specific rates, data residency and retention terms were not independently verified. Future open-source plans are not released artifacts. Biological and wet-lab demonstrations are not evaluated or covered.
Disclosures
AccessAllGPT received no payment, credits or private access for this briefing. Qwen, Alibaba and OpenRouter did not review or endorse it. AccessAllGPT Research is operated by NeuralArc and is not affiliated with these vendors. Publication-wide relationships are listed on the disclosures page.



