Dots vs Qwen 3.8 Max: The Difference Is Where You Are Allowed to Move It

Qwen 3.8 Max was published on August 3, 2026 and is Aliba​ba's highest-capability tier to date: natively multimodal, accepting text, images and video and returning text, across a 1,000,000-token context window, served through three wire formats — Open​AI-compatible, Anthro​pic-compatible and native DashScope — from Beijing, Singapore, Tokyo, Frankfurt and Virginia, at $2.00 per million input tokens and $6.00 per million output tokens with cached reads at $0.25. Dots is the company's always-on agent, announced at DevDay on September 29, 2026: it gets its own cloud computer and browser, reaches more than 4,000 apps, runs on GPT-6 Astra, and is included with Pro and Business Premium. Put side by side, the interesting question is not which is more capable. It is how many places each one is allowed to be, and who decides.

That is not an abstract distinction. It is the one that decides whether a tool is available in your market at all, whether it can be audited by your compliance team, and whether the work you build on it survives a change of terms. Capability is the easy part of this comparison; portability is the part that gets settled late and expensively.

Three ways to call one model

A wire format sounds like a small engineering detail until you have to move an application between vendors. The OpenAI-compatible surface means existing client libraries work without a translation layer. The Anthropic-compatible one means a codebase written against a different vendor's message shape does not have to be rewritten. The native DashScope interface exposes the full feature set, including the thinking toggle. Three surfaces for one model means the cost of switching is a base URL and a key, not a refactor.

That matters most when it is absent. Nothing about a dot offers a format, a URL or a key, because a dot is not something you call — it is something you talk to inside somebody else's interface. Moving your agent logic somewhere else means rebuilding it, then reconnecting the four thousand connectors it depended on, then discovering which of them behaved in ways you had not written down.

• What you are buying — a hosted worker inside ChatGPT, Slack and Teams (Dots) against a model endpoint you can call from anywhere (Qwen 3.8 Max)

• Interfaces — none published for a dot, against OpenAI-compatible, Anthropic-compatible and native DashScope surfaces (Qwen 3.8 Max)

• Where it runs — OpenAI's cloud computer with its own browser (Dots) against five served regions across Asia, Europe and North America (Qwen 3.8 Max)

• Reasoning control — not exposed for a dot, against an enable_thinking toggle, or reasoning.effort on the Responses API (Qwen 3.8 Max)

• Cost per unit — the first dot is included with Pro or Business Premium, allowance unpublished (Dots) against $2.00 per million input and $6.00 per million output, cached reads $0.25 (Qwen 3.8 Max)

• Independent evidence — vendor demos and capability statements, no benchmark for a dot, against an Artificial Analysis Intelligence Index of 45.4 at fifteenth of 145 models measured and 92.8 on GPQA Diamond (Qwen 3.8 Max)

Where the model lands, in numbers you can check

Qwen 3.8 Max's published profile is a mid-frontier one, and it is worth reading precisely rather than impressionistically. An AA Intelligence Index of 45.4 placing it fifteenth of 145 models measured puts it inside the top ten percent without putting it at the top. An AA Coding Index of 76.2 at ninth of 138 is stronger, and 88.8 on Terminal-Bench 2.1 is a genuinely high agentic-execution figure — higher than GPT-5.6 Sol's 88.0 on the same board. Where it is comparatively weaker is long-context recall at 80.3, which sits below Kimi K3's 88.7 and MiniMax M3's 83 despite a comparable window size, and SciCode at 52.1.

Alibaba's own positioning, in its migration guide, places the model directly against GPT-5.5, Claude Opus 4.7 and Gemini 3.1 Pro, and the honest reading of the numbers is that this is a model you reach for when a mix of strong agentic execution and multimodal input matters more than being first on the intelligence board. All of those figures are third-party measurements carried with their sources in our catalogue rather than vendor claims — which is the reason they are worth more than a launch deck.

One number is missing and should be said out loud rather than filled in: our catalogue does not list a maximum output length for this model. A million-token context is documented; a per-response output ceiling is not.

What Dots does not tell you, and why it is a design choice

OpenAI has published one sentence about what a dot costs and it does not contain a number: the first dot is included with Pro and Business Premium, with extended limits in the first month and no drawdown of ChatGPT usage limits for conversations with it. Beyond that there is no allowance figure, no second-dot price, no rate for extra capacity, and no per-task cost. The $500-per-month Pro 500 tier is a usage allowance plus the Ultrafast mode rather than a per-dot rate, so even the priciest plan does not convert into a cost per unit of agent work.

It is tempting to read that as a documentation gap. It is more accurate to read it as the category: a seat-priced product cannot have a per-unit price, because the whole point of a seat is that the marginal cost of using it is invisible. That is genuinely convenient for exploratory work and genuinely inconvenient for anything you have to attribute to a project, a client or a budget line.

Notice what the two products are optimised for. Qwen 3.8 Max is optimised for being called from anywhere, by anything, in whatever format the caller already speaks. A dot is optimised for being used in one place, by one interface, with the plumbing hidden. Both are coherent designs; they just answer different questions, and one of them answers "can I move this later" with a no.

Two answers to "what if this vendor changes"

The portability question gets concrete the moment a vendor moves a price, deprecates an identifier or changes what a subscription includes. For a metered model with three wire formats, the response is mechanical: point the client at a different base URL, or change one model string, and the application keeps working. There is nothing else to port because the model was never anything but a call.

For a dot, the response to a change in terms is to stop. The computer, the browser, the connectors and the allowance all belong to OpenAI, and none of them can be exported. That is not a criticism of the product; it is the honest description of renting. But it does mean that work you build entirely inside a dot has a migration path that runs through a rewrite, and that is a fact worth pricing in before you build something you intend to depend on for a year.

OrcaRouter is built for the first case. Qwen 3.8 Max is routed as qwen/qwen3.8-max at Alibaba's provider list price passed through with 0% markup, in the same catalogue as 200-plus other models behind one key, with automatic failover across upstream providers and a routing DSL for composing several models into a single request. That is what makes the three-region question answerable on your side rather than the vendor's: the same identifier reaches the model wherever it is served, and a rate change from Alibaba lands on your bill the same day instead of at the next renewal. What it does not reach is anything inside a dot — there is no endpoint for one, no identifier, and no substitute intelligence to swap in.

The question to ask before you commit

Ask what happens to your work on the day the vendor changes something you cannot control. If the answer is "we change a string and move on", you are buying a model, and the metered one is the better purchase. If the answer is "we would have to rebuild it", you are buying a product, and you should buy it deliberately — knowing that the convenience you are paying for is paid for in reversibility.

Neither answer is wrong. Buying portability you will never use is as wasteful as buying convenience you will need to escape. The mistake is not knowing which one you bought.