Verdict: Qwen 3.8 Max — Alibaba's 2.4-trillion-parameter flagship model — is the strongest signal yet that AI is shifting from a question-answering tool into an autonomous worker. It scored 86.1% on OSWorld-Verified, the benchmark that tests whether an AI can operate a real computer screen (clicking, typing, navigating apps), beating GPT-5.6 Sol Max (83.2%), Gemini 3.1 Pro (76.2%), and Claude Fable 5 (85.0%) — all vendor-reported scores from Qwen's launch benchmark table. For businesses, the practical shift is not a bigger brain but fewer tasks that need a human babysitting every prompt. The agent now runs for days on its own.
Last verified: 2026-08-08
- 2.4T total parameters, 95B active per token (MoE architecture)
- 86.1% on OSWorld-Verified — #1 on the public leaderboard as of August 7, 2026
- 5 built-in tools: code interpreter, web search, web extractor, text-to-image search, reverse image search
- 16-day autonomous coding demo produced 265 commits, 127 PRs, 151 issues — public GitHub repo
- $2 / $6 per million tokens (input / output) via QwenCloud API
- Open weights promised by Alibaba, expected "next week" after August 3 release
Pricing, limits, and model versions change often — last checked August 8, 2026.
What makes Qwen 3.8 Max different from other large language models?
Most large language models are built to answer one prompt at a time. Qwen 3.8 Max is built to plan, use tools, carry context across a million tokens, and keep running — for hours or days — without human intervention at every step. The distinction matters for any team evaluating where to route real production work. Most benchmark scores you see online measure single-shot quality. Qwen 3.8 Max's headline scores measure long-horizon task completion, the kind of work that ends in a finished deliverable rather than a chat answer.
The architecture is a sparse Mixture-of-Experts (MoE) model with 2.4 trillion total parameters, of which approximately 95 billion activate per token — confirmed by Alibaba's official release announcement and independently tracked by AI Release Tracker. This means the full parameter count does not fire on every request; the model routes each token through the expert subset best suited to it, which keeps inference cost practical despite the headline size. The context window is 1 million tokens, which is roughly equivalent to processing a large codebase or a multi-thousand-page document set in one pass.
How does Qwen 3.8 Max score on computer-use benchmarks?
On OSWorld-Verified — the benchmark that evaluates whether an AI agent can operate a real computer desktop environment by clicking buttons, typing text, opening applications, and completing multi-step tasks — Qwen 3.8 Max scored 86.1%, the highest published score of any tracked model as of August 7, 2026, per the BenchLM.ai leaderboard.
Here is how the top models compare on OSWorld-Verified:
| Model | Developer | OSWorld-Verified Score | Context Window | Data Source |
|---|---|---|---|---|
| Qwen 3.8 Max | Alibaba / Qwen | 86.1% | 1M | Qwen launch benchmarks |
| Claude Fable 5 | Anthropic | 85.0% | 1M+ | Qwen launch benchmarks |
| Claude Opus 4.8 | Anthropic | 83.4% | — | Qwen launch benchmarks |
| GPT-5.6 Sol Max | OpenAI | 83.2% | — | Qwen launch benchmarks |
| Gemini 3.1 Pro | Google DeepMind | 76.2% | 2M | Qwen launch benchmarks |
Important caveat: All scores in this table are from Qwen's own launch benchmark comparison, published alongside the model's August 3, 2026 release. They have not been independently verified by a third party at the time of writing. The operator of the OSWorld-Verified leaderboard lists all 22 tracked models as "self-reported." Treat these as vendor-reported until independent evaluations arrive.
What is the "Agent OS" concept and why does it matter?
The phrase "agent operating system" gets used loosely, but the idea is concrete: a model plus a set of built-in tools that let it act on the environment rather than just generate text. Think of it as the difference between a brain in a jar and a brain with hands.
Qwen 3.8 Max ships with five built-in tools accessible through its Responses API, confirmed in QwenCloud's official documentation:
| Tool | What it does | Practical use |
|---|---|---|
| Code interpreter | Writes and runs Python code in a sandbox | Data analysis, math, testing iterations |
| Web search | Retrieves live information from the internet | Real-time research, competitor monitoring |
| Web extractor (scraping) | Pulls actual text content from a specified URL | Reading full pages instead of guessing from snippets |
| Text-to-image search | Finds relevant images based on a text description | Visual research, product comparisons |
| Reverse image search | Finds visually similar images from an input image | Finding similar products, visual content tracing |
Before these built-in tools, you had to hand-roll a web-search wrapper, a code sandbox connection, and an image matching pipeline — then wire each one into your model's function-calling interface. Now they are a server-side configuration flag. The model decides which tool to invoke based on the question, with no additional setup required.
For businesses already building agentic operating systems around other models, this matters: the tool layer is part of the model, not something you maintain separately. If you are curious how this architecture compares to a full Claude-based agent OS, the distinction is that Claude's built-in tool use requires external harnessing while Qwen 3.8 Max's tools are native to the API.
How did Qwen 3.8 Max perform on a 16-day autonomous coding run?
Alibaba's team started Qwen 3.8 Max on an empty coding project with no human hand-holding and let it run autonomously for approximately 16 days. By July 30, 2026, the project — a CLI tool called "oh-my-cli" — had accumulated 265 commits, 127 pull requests, and 151 issues. The repository is public on GitHub under the Apache 2.0 license, meaning every commit, PR, and issue timestamp is auditable. This was confirmed by DataCamp's independent coverage and The New Stack.
Here is where it gets honest. An independent audit by THE DAILY BRIEF opened the actual GitHub repository and found:
- The repository has two contributors: the bot account
qwen-code-dev-bot(635 commits as of August 5) and a humanqqqys(13 commits). That is approximately 98% machine-authored. - Alibaba's own announcement describes the run as "more than ten days," not strictly 16 continuous uninterrupted days. The 16 days is the elapsed window.
- The human contributor built the autonomy governance contract before the agent was allowed to run, and fixed the CI workflow on the morning of the announcement.
- The repository has continued growing past the snapshot: as of August 5 it held 648 commits, not 265.
This is not a criticism — it is the most transparent autonomous coding demonstration any major lab has published. But it means the headline should read "an autonomous coding loop with a human-built governance layer ran for over ten days and produced hundreds of commits," not "an AI coded for 16 days with zero human involvement."
For a broader look at how autonomous coding agents have evolved, see our analysis of the five levels of AI coding adoption — Qwen 3.8 Max sits between Level 4 (autonomous with human governance) and Level 5 (fully self-directed).
Can Qwen 3.8 Max reproduce and improve research papers?
Yes, and this is one of its most distinctive capabilities. Alibaba's team gave Qwen 3.8 Max a recent machine learning research paper and asked it to reproduce the experiments in code from scratch — no starter code, no ready-made pipeline — then try to beat the paper's own method. According to the Qwen blog post and verified by DataCamp:
- The model worked for approximately 5 days (~125 hours of continuous effort).
- It wrote roughly 7,600 lines of code and took over 1,100 actions.
- It ran 33 rounds of GPU training.
- It first spent ~37 hours rebuilding the paper's full pipeline and reproduced all six main findings.
- Then it ran a self-improving research loop across 4 rounds, testing 18 improvement ideas.
- Its final invented method beat the paper's original approach by 2.71 points on AIME24, a competition-level math benchmark.
This is the kind of work that normally requires a research team's week of setup before any experiments begin. Whether it generalizes beyond this specific paper is unknown — it is a single vendor-demonstrated case. But the signal it sends to teams evaluating model capability for long-horizon tasks is that the context window, tool use, and feedback loop are all configured to sustain multi-day work without restarting.
Where does Qwen 3.8 Max fall short?
Qwen 3.8 Max is not a universal winner. Its honest weak spot is production-grade software engineering, measured by benchmarks that test resolving real GitHub issues in professional codebases.
| Benchmark | What it tests | Qwen 3.8 Max | Claude Fable 5 | Source |
|---|---|---|---|---|
| SWE-bench Pro | Resolving real GitHub issues in professional codebases | 67.7% | 80.0% | Qwen launch benchmarks |
| Terminal-Bench 2.1 | Agent working in a terminal to complete engineering tasks | 86.6% | 84.6% | Qwen launch benchmarks |
| PaperBench | Reproducing a research paper's experiments in code | 93.0% | 88.8% | Qwen launch benchmarks |
| HLE (Humanity's Last Exam) | Extremely hard reasoning tasks | 43.6% | 53.3% | Qwen launch benchmarks |
If your primary use case is writing production-level software all day against a complex monorepo, the 12-point gap on SWE-bench Pro behind Claude Fable 5 is real and matters. Alibaba does not claim to beat every model at everything; they claim a model that is strong across a huge range of jobs — coding, research, computer use, images — instead of being a specialist in one lane. The scores back that claim up more than they don't.
For a detailed task-by-task routing guide comparing Qwen 3.8 Max to other frontier models, read our companion analysis: Qwen 3.8-Max vs Claude Fable 5 vs GPT-5.6 Sol: which model wins your next task?
How much does Qwen 3.8 Max cost and how do you access it?
API pricing: $2.00 per million input tokens and $6.00 per million output tokens, with cached input at $0.25 per million tokens, confirmed by OpenRouter and TokenCost. For comparison:
| Model | Input / 1M tokens | Output / 1M tokens | Context |
|---|---|---|---|
| Qwen 3.8 Max | $2.00 | $6.00 | 1M tokens |
| Claude Fable 5 | ~$10.00 | ~$50.00 | 1M+ tokens |
| GPT-5.6 Sol Max | — | — | — |
| Kimi K3 | ~$3.00 | ~$15.00 | ~1M tokens |
Qwen 3.8 Max undercuts Claude Fable 5 significantly on output cost — roughly 8x cheaper per output token — which matters if your workload is long-form generation or long-horizon agent loops that produce many tokens.
Access routes:
- QwenCloud API (Alibaba's hosted platform) — OpenAI-compatible and Anthropic-protocol-compatible, so integration is primarily a base-URL and model-ID change (Qwen blog)
- OpenRouter — model ID
qwen/qwen3.8-max, live since day one (OpenRouter model page) - Vercel AI Gateway — added as
alibaba/qwen3.8-maxat provider pricing - Tools including Claude Code, Cursor, Cline, Codex, and OpenClaw can point at it through existing integrations since the API speaks the same protocols they already use
Is Qwen 3.8 Max open source?
No — not yet. As of August 8, 2026, Qwen 3.8 Max is a proprietary model available only through Alibaba's own API and partner platforms. However, Alibaba announced at release that it intends to publish the open weights — which would make it by a wide margin the largest model any lab has open-sourced. A smaller Qwen3.8-27B checkpoint is also slated for open-weight release. Both are expected on Hugging Face and ModelScope "next week" after the August 3 launch, per the Qwen blog and DataCamp.
Until the weights actually ship, Qwen 3.8 Max is a hosted API with a roadmap, not a self-hostable model. If running a private version on your own infrastructure is the deciding factor, the current open frontier includes models like open-weight agentic tools and local agentic OS setups that already have published weights.
Should you route business data through a China-based AI platform?
This is a question worth answering directly. Qwen 3.8 Max runs through Alibaba's own platforms, including QwenCloud. Routing business data through a platform based in another country requires the same due diligence you would apply before connecting any new tool to your data:
- Where is your data physically processed and stored?
- What is the provider's data retention policy?
- Does the provider train on customer inputs by default, and can you opt out?
- Are there export-control or data-sovereignty regulations in your industry that restrict cross-border data flows?
This is not a reason to avoid the model. It is a reason to classify your data before you connect. Non-sensitive research, public-facing content, and general-purpose coding tasks are low-risk. Customer PII, proprietary source code, and regulated data require a data-processing agreement or, after the open weights ship, a fully self-hosted deployment. If you are already running autonomous AI agents with shared memory in your own infrastructure, you have the architecture to swap in a self-hosted model once the weights arrive.
What does Qwen 3.8 Max mean for your business?
For small businesses and operations teams: AI is moving from a tool you type into, to a worker you delegate to. The practical shift is not about parameter counts — it is about fewer tasks that need a human standing over the agent's shoulder at every step. Start planning for workflows where an agent runs a multi-step research, content, or coding task for hours, then hands you a finished result with an audit trail.
For developers and builders: Qwen 3.8 Max's built-in tools and OpenAI-compatible API mean you can test it by changing a base URL in your existing agent harness. Run one demanding task with a clear success condition, record the total tokens and retry count, and compare the deliverable — not just the first answer. When the open weights ship, you will be positioned to swap from a hosted API to a self-hosted deployment without re-architecting your stack.
For teams evaluating cloud vs. self-hosted agents: The open weights announcement is the signal to start planning. Running a 2.4-trillion-parameter model locally requires serious GPU infrastructure, but the smaller 27B checkpoint is designed for ordinary on-premise hardware. If your business depends on data sovereignty or cost predictability, the 27B model on your own infrastructure may be the more durable play than renting API access from any vendor.
FAQ
Q: What is Qwen 3.8 Max? A: Qwen 3.8 Max is Alibaba's flagship AI model released on August 3, 2026. It has 2.4 trillion total parameters (95 billion active per token via Mixture-of-Experts), a 1-million-token context window, and five built-in tools (code interpreter, web search, web extractor, text-to-image search, reverse image search). It is designed for long-horizon agentic tasks rather than single-shot question answering.
Q: How does Qwen 3.8 Max compare to Claude Fable 5? A: On computer-use (OSWorld-Verified), Qwen 3.8 Max scored 86.1% vs Claude Fable 5's 85.0% — but these are vendor-reported scores. On software engineering (SWE-bench Pro), Claude Fable 5 leads at 80.0% vs Qwen's 67.7%. Qwen 3.8 Max is roughly 8x cheaper per output token ($6 vs ~$50 per 1M tokens). Qwen excels at broad agentic tasks; Claude Fable 5 excels at production software engineering. All scores are vendor-reported from Qwen's launch benchmarks.
Q: Can Qwen 3.8 Max really code autonomously for 16 days? A: Partly. Alibaba's 16-day autonomous coding run produced 265 commits, 127 PRs, and 151 issues in a public GitHub repository. An independent audit confirmed the result is approximately 98% machine-authored, but a human contributor built the governance contract before the agent started and contributed 13 of 648 total commits. Alibaba's own announcement describes the run as "more than ten days" of continuous operation, not strictly 16 uninterrupted days.
Q: Is Qwen 3.8 Max open source? A: No, not yet. As of August 8, 2026, Qwen 3.8 Max is available only through Alibaba's hosted API (QwenCloud) and partner platforms like OpenRouter. Alibaba has announced that open weights for both Qwen3.8-Max and a smaller Qwen3.8-27B will be released on Hugging Face and ModelScope, expected within a week of the August 3 launch.
Q: How much does Qwen 3.8 Max cost? A: Qwen 3.8 Max costs $2.00 per million input tokens and $6.00 per million output tokens via QwenCloud, according to OpenRouter and TokenCost as of August 2026. Cached input tokens cost $0.25 per million. This makes it significantly cheaper than Claude Fable 5 (~$10/$50 per million tokens) for output-heavy workloads.
Q: Is it safe to route business data through Alibaba's Qwen platform? A: It depends on your data classification. Non-sensitive research, public content, and general coding tasks are low-risk. For customer PII, proprietary code, or regulated data, you should review Alibaba's data processing terms, opt out of training on your inputs if available, or wait for the open weights to self-host. The same data-sovereignty due diligence applies to any cloud-based AI platform.
Every claim here is traced to a primary source, dated, and listed under Sources. Research and drafting are AI-assisted; editing, verification and publication are human decisions, and a person is accountable for what appears on this page. How we work →
Discussion
0 comments