A closer look at Qwen4-Max
Qwen4-Max gives teams another model to test for the work that happens between a brief and a finished deliverable. The launch examples include spreadsheets, presentations, game development and long research assignments. The detail that stands out is the combination of better reported results and lower task costs. On Qwen’s internal Work360 evaluation, the score rises from 47.74 to 52.69 while average cost per task falls from $4.43 to $1.67.
If you run a small team or build things on your own, that is a useful question to investigate: can this model handle a worthwhile part of your work at a lower cost, without creating more review and correction? The launch gives us evidence to start testing. It does not answer that question for every workflow.
Our Qwen4-Max model page brings together the published results and API specifications. It does not yet have a complete Agent Index: 4 of the seven required components have matching results. LiveBench Agentic, Coding and Browser Use are still missing. We checked the current LiveBench board and approved feed on November 16; neither lists Qwen4-Max. The updater is registered to pick up its matching results when they are published. Missing benchmarks remain blank, so we do not assign it a composite rank.
Choosing specifically between these two? Read the Qwen4-Max vs Fable 5.2 decision report for every shared benchmark and a closer look at workload cost.
Qwen4-Max, Fable 5.2 and OpenAI compared
Switch between seven independent benchmark tracks to compare Qwen4-Max with Claude Fable 5.2 and OpenAI’s current GPT-6.1 Sol, GPT-6.1 Terra and GPT-6 Astra. Each chart keeps the benchmark version and evaluation setup aligned. Effort labels and linked scores show which reported configuration you’re comparing.
Same benchmark version and evaluation track. Reasoning effort differs and is labelled. The latest dated report is used, preferring the highest available effort on that date. Missing reports stay blank.
All benchmark scores and sources
These tables contain every comparison shown in the chart selector. Download all guide chart data as JSON ↗
HLE-Diamond · 1,000 questions · no tools
| Model | Effort | Score | Source |
|---|---|---|---|
| Qwen4-Max | xhigh | 44.9% | Published result ↗ |
| Claude Fable 5.2 | max | 80.2% | Published result ↗ |
| GPT-6.1 Sol | max | 58.8% | Published result ↗ |
| GPT-6.1 Terra | max | 47.3% | Published result ↗ |
| GPT-6 Astra | max | 63.8% | Published result ↗ |
GDP.pdf · all-pass · five attempts
| Model | Effort | Score | Source |
|---|---|---|---|
| Qwen4-Max | xhigh | 26.2% | Published result ↗ |
| Claude Fable 5.2 | max | 24.2% | Published result ↗ |
| GPT-6.1 Sol | max | 31% | Published result ↗ |
| GPT-6.1 Terra | max | 28.6% | Published result ↗ |
| GPT-6 Astra | max | 31% | Published result ↗ |
Terminal-Bench 4.0 · mini-swe-agent · three repeats
| Model | Effort | Score | Source |
|---|---|---|---|
| Qwen4-Max | xhigh | 48.48% | Published result ↗ |
| Claude Fable 5.2 | max | 59.6% | Published result ↗ |
| GPT-6.1 Sol | max | 56.06% | Published result ↗ |
| GPT-6.1 Terra | max | 55.6% | Published result ↗ |
| GPT-6 Astra | xhigh | 59.6% | Published result ↗ |
FrontierCode 2.0 · mini-swe-agent · three repeats
| Model | Effort | Score | Source |
|---|---|---|---|
| Qwen4-Max | xhigh | 26.45% | Published result ↗ |
| Claude Fable 5.2 | max | 48.76% | Published result ↗ |
| GPT-6.1 Sol | max | 40.63% | Published result ↗ |
| GPT-6.1 Terra | max | 42.32% | Published result ↗ |
| GPT-6 Astra | max | 42.37% | Published result ↗ |
Terminal-Bench-Science 0.1 · independent mini-swe-agent
| Model | Effort | Score | Source |
|---|---|---|---|
| Qwen4-Max | xhigh | 23.33% | Published result ↗ |
| Claude Fable 5.2 | max | 73.33% | Published result ↗ |
| GPT-6.1 Sol | max | 58.1% | Published result ↗ |
| GPT-6.1 Terra | max | 57.62% | Published result ↗ |
| GPT-6 Astra | max | 63.33% | Published result ↗ |
AutomationBench-AA · v1.0.6 · objective completion
| Model | Effort | Score | Source |
|---|---|---|---|
| Qwen4-Max | xhigh | 72.56% | Published result ↗ |
| Claude Fable 5.2 | max | 76.42% | Published result ↗ |
| GPT-6.1 Sol | xhigh | 66.6% | Published result ↗ |
| GPT-6.1 Terra | max | 60.2% | Published result ↗ |
| GPT-6 Astra | max | 68.5% | Published result ↗ |
AA-Briefcase v1.1 · Stirrup · combined Elo
| Model | Effort | Score | Source |
|---|---|---|---|
| Qwen4-Max | xhigh | 1,686.07 Elo | Published result ↗ |
| Claude Fable 5.2 | max | 1,785.32 Elo | Published result ↗ |
| GPT-6.1 Sol | max | 1,555 Elo | Published result ↗ |
| GPT-6.1 Terra | max | 1,402 Elo | Published result ↗ |
| GPT-6 Astra | max | 1,569 Elo | Published result ↗ |
The launch-only Work360, Agents’ Last Exam, DeepSWE and Chartography figures below keep their own protocols. They are not substituted into these independent comparisons.
Benchmarks explained
Qwen improves over its predecessor across many launch tests, but its relative strengths vary. Here are six results that help frame where to try it. The first comparison bars retain the launch snapshot; independent results are identified separately.
Work360 is Qwen’s internal evaluation. 52.69 is its reported score; the launch does not establish that 52.69% of arbitrary client projects will be completed successfully.
View published source ↗Long jobs need checkpoints you can review.
The 18.8-point increase on this long-horizon task evaluation is one of the larger changes in the launch chart. It makes multi-step assignments worth testing: review a folder, identify the important issues, produce a document, then check it against the brief. Qwen’s Live Delivery approach returns usable versions along the way so the person directing the work can correct the course early.
Launch-chart scores. The exact task release, agent harness, effort and time budget are not specified in the article’s footnotes. Kept separate from Google’s October-task-set track.
View published source ↗Coding gains are uneven across tests.
Qwen4-Max improves by 6.4 points on DeepSWE. That is encouraging for repository work, but the same launch chart puts it at 31.5 on ProgramBench, compared with 65.0 for Opus 5.5. The practical move is to test it on the kind of engineering your agent actually does: fixing bugs, implementing features or reproducing an existing application. A strong result in one category does not settle the others.
Qwen’s launch comparison. DeepSWE and ProgramBench measure different tasks; these figures are not developer productivity estimates. The prior-model label in the chart is Qwen3.8-Max-0902.
View published source ↗Useful workflow progress still needs an end-to-end check.
Artificial Analysis reports about 72.6% on objective-level workflow scoring at xhigh. For a business, that suggests testing sequences such as finding a record, applying a policy and preparing the next step. A partly correct workflow can still leave the customer waiting, so your own evaluation should check whether the entire job finished and whether any actions were wrong.
Same benchmark version and evaluation track. Reasoning effort differs and is labelled. The latest dated report is used, preferring the highest available effort on that date. Missing reports stay blank.
Judge the files it hands back.
AA-Briefcase evaluates complex knowledge work and the quality of the resulting deliverables. Qwen4-Max’s combined rating is 1686.07 Elo in the independent record. This is relevant when your output is a workbook or presentation someone else must use. Give it source files and a clear rubric, then inspect the calculations, analytical quality and presentation before handing the work to a client.
Same benchmark version and evaluation track. Reasoning effort differs and is labelled. The latest dated report is used, preferring the highest available effort on that date. Missing reports stay blank.
Native vision helps with documents, but check the numbers.
The launch reports a 12.2-point improvement in chart understanding. That is useful for work that starts with a deck or a visual report: read a chart, identify a trend and turn it into an explanation. Keep the original data nearby. A chart-reading benchmark does not prove the model will reliably extract every value from a client’s slide.
Qwen sets min_pixels to 4096×32×32. This is a separate protocol from the Anthropic implementation and its grader. Tool access is unspecified.
View published source ↗What useful work costs
QwenCloud lists standard Singapore pricing at $1.50 per million input tokens and $4.50 per million output tokens. Implicit cached input costs $0.075 per million tokens. A request using 100,000 uncached input tokens and 20,000 output tokens costs $0.24 for those tokens. Tool charges, retries and any further generation add to that bill.
| Internal evaluation | Qwen3.8-Max | Qwen4-Max |
|---|---|---|
| Work360 score | 47.74 | 52.69 |
| Average cost per task | $4.43 | $1.67 |
| 1,000 similar attempts | $4,430 | $1,670 |
The reported average falls about 62%, a $2,760 difference over 1,000 similar attempts. These are Qwen’s internal evaluation costs per attempted task. They do not measure cost per successfully completed client project. Source: Qwen launch.
The important change for an agency could be the time spent making the output useful. A low token bill is attractive until someone spends an hour fixing a spreadsheet. Track both the model spend and the human review time when you test it.
Blended token cost
USD per 1M tokens · 50% input + 50% output · lower is better
Pricing details and sources
Blended cost = (input price + output price) ÷ 2: 500,000 input and 500,000 output tokens. Each model uses its lowest sourced current provider rate, excluding batch and Flex tiers. Cache discounts, tool fees and workload-specific token usage are excluded. 52 models without complete sourced current prices are unavailable in this comparison. The default selection is this model and the nine closest in blended cost; if this model has no price, we show the ten cheapest. The scale stays fixed across pages of your selection.
| Model | Input / 1M | Output / 1M | Blended / 1M | Provider / source |
|---|---|---|---|---|
| GLM-5.2 | $1.40 | $4.40 | $2.90 | Z.aiGLM-5.2 (Max) · Checked 2026-11-02 |
| GPT-6.1 Terra | $1.00 | $5.00 | $3.00 | OpenAI ↗Standard API, prompts up to 272K input tokens · Checked 2026-11-02 |
| Qwen4-Max | $1.50 | $4.50 | $3.00 | QwenCloud ↗Standard API, Singapore; uncached input · Checked 2026-11-16 |
| Gemini 4 Hydrogen | $1.50 | $7.50 | $4.50 | Google ↗Standard Gemini API; independent speed measured at xhigh · Checked 2026-11-12 |
| Claude Sonnet 5.5 | $2.00 | $10.00 | $6.00 | AmazonClaude Sonnet 5.5 (max with fallback) · Checked 2026-11-02 |
| GPT-6.1 Sol | $2.00 | $10.00 | $6.00 | AzureGPT-6.1 Sol (max) · Checked 2026-11-02 |
| Kimi K3.1 Pro | $2.00 | $10.00 | $6.00 | KimiKimi K3.1 Pro · Checked 2026-11-02 |
| Claude Opus 5.5 | $4.00 | $20.00 | $12.00 | AmazonClaude Opus 5.5 (max with fallback) · Checked 2026-11-02 |
| Claude Fable 5.2 | $8.00 | $40.00 | $24.00 | Anthropic ↗Standard uncached API · Checked 2026-11-09 |
| GPT-6 Astra | $10.00 | $50.00 | $30.00 | AmazonGPT-6 Astra (max) · Checked 2026-11-02 |
This chart uses an equal input/output token blend. Your mix may differ substantially, especially for agents that repeatedly read a growing context. Cache reuse and token efficiency can change the cost of the whole run.
Coding agents and personal projects
Qwen reports using roughly 48% fewer tokens at Xhigh and 49% fewer at High on Terminal-Bench 4.0 compared with Qwen3.8-Max. That is a token-efficiency result in its setup. It does not establish that an agent finishes twice as quickly.
Independent testing gives a more specific view. Artificial Analysis’s xhigh record reports 48.48% on Terminal-Bench 4.0, 26.45% on FrontierCode 2.0 and 23.33% on Terminal-Bench-Science. Those evaluations use their own recorded protocols. You can explore them on the model page alongside the launch results.
For your coding agent, try a small set of real jobs with known acceptance checks: fix a reproducible bug, add a feature to an existing application and clean up a failing test. Measure whether the changes work and how much you need to intervene. The launch’s weaker ProgramBench result is a reason to include a demanding implementation task before moving a whole workflow over.
For personal projects, native vision and tool calling open up more than chat. Qwen demonstrates games built with Unity, Unreal Engine and Godot, and scenes created in Blender. Treat those as selected demonstrations of what the model-plus-tools setup can produce. Start with a bounded project you can run and inspect, then add complexity after the first version works.
Speed and context
Qwen4-Max supports a published one-million-token context window with text, image and video inputs and text output. A large context can help you work across repositories or source-document collections, but fitting material into a request does not guarantee every detail will be used correctly.
The released-model API documentation lists 131,072 output tokens without thinking, or 262,144 combined thinking and final-answer tokens. The catalog records the combined budget and preserves this distinction in its source notes. These exact limits conflict with the marketplace’s rounded 393K label; we use the model-specific API documentation. Qwen4-Max Preview is a separate model ID. A deeper November 16 check of Artificial Analysis’s model and provider records found no output-speed, time-to-first-token or first-answer measurements for Qwen4-Max. Its knowledge cutoff is also undisclosed in the reviewed official sources.
For a long agent run, the useful timing measure is how long it takes to return an acceptable result, including tools and corrections. Qwen’s Live Delivery plugin is designed to return usable stages early. That gives you a chance to correct a misunderstood brief before the agent spends more time on it.
Access and open weights
Qwen4-Max is available through QwenCloud’s API, Qwen Studio, QwenWork and Qoder. The API offers tool calling and structured outputs, with OpenAI- and Anthropic-compatible interfaces.
Qwen4-Max is a proprietary release. Qwen says smaller Qwen4 models will receive open weights later. That promise does not make Max an open-weight model or establish a license for the future releases. If you need self-hosting today, use the leaderboard’s open-model view to research available alternatives.
Sources and comparison limits
All guide chart data, methodology and sources as JSON ↗
This report uses Qwen’s launch article and benchmark chart, its QwenCloud marketplace page, and the independent Artificial Analysis xhigh record checked on November 16, 2026. Qwen’s article is visibly dated November 15; the linked X announcement was posted November 16 UTC.
The launch reports 51.7 on Humanity’s Last Exam without specifying its subset or tool access. The independent record reports 44.9% on the separate, tool-free HLE-Diamond evaluation. Those scores belong to different tests and remain separate. The same care applies to launch-chart results versus independent agent harnesses, and to rounded Elo ratings versus pass percentages.
The question I would take into a trial is simple: which part of a real job can Qwen4-Max finish well enough to reduce your total cost? For a small team, that may be a first spreadsheet, a research pass or a coding task with clear tests. Give it a concrete brief, inspect the result and compare the total effort with the model you already use.

A cheaper first pass at professional work.
Think about a client brief that needs a spreadsheet, a written analysis and a presentation. Qwen’s Work360 framework covers 26 broad industry domains, with examples ranging from campaign reviews to engineering handoffs. Its reported score rises by 4.95 points over Qwen3.8-Max. That gives a small team another model to test on the production work around a project: gathering evidence, reconciling files and building a first deliverable.