Treat the 35% token saving as a planning hypothesis, not a forecast for your agent bill. Vercel’s 15 November 2026 announcement makes Qwen4-Max available through AI Gateway; that is gateway availability, not evidence of native provider support in Claude Code, Codex, Cursor or another agent framework. Alibaba reports 35% fewer tokens than Qwen3.8-Max on Work360 at High reasoning effort, but its own Qwen4-Max post reports approximately 24% at Xhigh. For a draft budget, use the figure matching the effort you will actually run, then replace it with your own cost per accepted coding task.
That distinction changes the decision. A team running Xhigh should not build its business case around 35%. Nor should it credit a token reduction until failed attempts, automatic retries, cached input and output charges have been included. The supplied evidence contains no independent replication of Alibaba’s token or cost results, no published Work360 task list or scoring rubric in the fetched text, and no Qwen4-Max per-token rate in the fetched Vercel pricing page. The useful next step is therefore a controlled comparison, not a switch based on the changelog.
What the gateway announcement changes
Vercel describes Qwen4-Max as a reasoning model for coding and long-running agent tasks, with image input, a vendor-stated 1M-token context window, tool calling and structured outputs. The AI Gateway model ID is alibaba/qwen4-max. Vercel’s documented setup procedure is to install the latest Vercel CLI, run npx vercel ai-gateway setup, follow the setup instructions and select that model ID in the agent.
Those steps establish a route to the model. They do not establish that an agent framework has a native Qwen provider, that every framework feature works identically through the gateway, or that tool calling and structured outputs behave the same way across integrations. Keep those as test items rather than compatibility assumptions.
Also keep the release surfaces separate. StealthModels reports a preview API model ID of qwen4-max-preview, a one-million-token context and up to 393,216 output tokens, with text, image and video input. Those preview details should not be transferred automatically to alibaba/qwen4-max on AI Gateway. Alibaba separately lists QwenCloud, through its Token Plan or API, along with QwenWork, Qoder and QwenStudio in its availability announcement. Different routes can have different limits, prices and release status; results from one are not automatically results from another.
The headline saving depends on reasoning effort
Alibaba’s reported reductions are not a single model-wide constant:
| Vendor-reported comparison with Qwen3.8-Max | Xhigh effort | High effort |
|---|---|---|
| Work360 token usage | Approximately 24% lower | Approximately 35% lower |
| Terminal-Bench 4.0 token usage | Approximately 48% lower | Approximately 49% lower |
Both rows come from Alibaba’s Qwen4-Max evaluation post. No fetched source independently replicates them.
The effort qualifier matters operationally because StealthModels reports that Qwen Code v0.26.5 gives the qwen4 model IDs the same reasoning levels as Qwen3.8-Max: low, medium and xhigh, with xhigh as the default. On that reporting, approximately 24% is the closer vendor-supplied planning figure for a team leaving Qwen Code at its default. It is still not a guaranteed saving.
Do not silently equate a benchmark label such as High with a similarly named setting in a client. Record the exact effort value sent with each request and confirm that both models receive it. A comparison between Qwen4-Max at a lower effort and Qwen3.8-Max at a higher effort could produce a token saving while confounding model changes with changes in the amount of reasoning requested.
The benchmark scope also limits what can be inferred. Alibaba describes Work360 using a reported taxonomy of 26 Tier-1 domains, 182 Tier-2 subdomains and 721 Tier-3 specialized fields, but the fetched text does not include the detailed tasks, scoring rubric or footnoted methodology. That taxonomy shows intended breadth; it does not show whether the evaluated work resembles your repositories, tool traces or acceptance criteria.
Build the comparison around completed tasks
The unit to compare is not price per million tokens or tokens per isolated response. It is the total cost of getting a fixed body of work to a predefined quality threshold.
A practical worksheet should have one row per attempt and one summary per model. The following fields make the comparison reproducible; they are an evaluation protocol derived from the reported conditions and platform features, not a vendor-supplied testing method.
| Worksheet field | What to record | Why it can change the decision |
|---|---|---|
| Model and route | Exact model ID, gateway or direct API, client and client version | Prevents preview, gateway and QwenCloud results from being pooled |
| Reasoning effort | Exact requested value for every attempt | Separates model effects from effort changes |
| Task set | Fixed task identifiers, repository state and acceptance criteria | Makes completed-task quality comparable |
| Tool surface | Tool list, schemas and permissions held constant | A changed tool interface can change traces and outcomes independently of the model |
| Context and output limits | Configured limits and any truncation or rejection | A large advertised context does not make two route configurations equivalent |
| Token accounting | Uncached input, cached input and output for every attempt | Token categories can have different prices |
| Retries and reruns | Attempt number, retry reason, responding model and tokens for each additional attempt | Failed attempts can consume the apparent saving |
| Completion quality | Pass or fail against the same predefined threshold for both models | Fewer tokens are not useful if accepted completions fall |
| Latency | Time per attempt and time to an accepted completion | The supplied sources provide no comparative latency result |
| Prices | Dated rates for each token category and route | The fetched pages do not supply a Qwen4-Max AI Gateway per-token rate |
| Release status | Endpoint status and preview or production terms on the test date | The supplied timeline is not consistent enough to assume durable availability |
Run the same tasks against both models with tools, reasoning effort and quality thresholds held constant. Keep retry policy constant as well. Then run a second configuration if your production service uses managed fallback or automatic retries, so you can see the economics of the configuration you would actually operate. These are recommended controls, not reported benchmark procedures.
For quality, use acceptance criteria appropriate to the work, such as the repository’s existing checks plus a consistent review standard. The important control is that a completion counts only when it meets the same bar for both models. Do not loosen the threshold for the model that uses fewer tokens.
Split the token bill by attempt
For each attempt, record uncached input tokens (U), cached-input tokens (K), output tokens (O), and the confirmed price for each category. A basic cost expression is:
[ C_a = U_a r_U + K_a r_K + O_a r_O ]
Sum (C_a) across the initial attempt, retries and human-requested reruns. Report retry counts and retry tokens separately as diagnostics, but do not add them to the total twice if their input and output tokens are already included in the per-attempt records.
Then calculate:
[ \text{Cost per accepted task} = \frac{\text{Cost of every attempt on the fixed task set}} {\text{Number of tasks meeting the quality threshold}} ]
This denominator is important. If one model produces fewer accepted tasks, the cost of its failed work does not disappear. If no tasks pass, there is no finite cost per accepted task to compare.
Cached input deserves its own column because Alibaba says it has lowered API prices relative to Qwen3.8-Max, with particularly large reductions in cached-input pricing. That makes the average Work360 cost result potentially sensitive to both token consumption and price mix. Your own caching pattern may differ, so measure cached input rather than assuming the benchmark’s balance.
Long tool traces create a similar issue. If a model saves reasoning tokens but requires more tool attempts, the additional inputs and outputs can offset the initial reduction. That is a mechanism to measure, not a reported Qwen4-Max failure rate: the supplied sources do not provide comparative tool-trace or retry data.
Vercel’s pricing page lists AI Gateway features including bring-your-own-key, managed fallback, load balancing, spend monitoring and automatic retries, along with 500 included token requests per month. Automatic retries are especially relevant to the worksheet because a request-level summary can hide additional model attempts. Log the responding model on every attempt as well; otherwise a fallback could turn a two-model comparison into an unlabelled mixture.
Price the run with rates you have actually confirmed
Alibaba reports that average cost per Work360 task fell from $4.43 to $1.67, a vendor-reported 62% reduction. That figure combines its evaluation, token usage and pricing assumptions. It is not an independently measured AI Gateway bill, and the fetched Work360 methodology is insufficient to reconstruct it.
The fetched Vercel pricing page does not show a Qwen4-Max per-token rate in its visible text. Before running the comparison, record the current rates for the exact route you intend to use and the date you checked them. If you cannot confirm a rate, leave the corresponding cost result unresolved rather than borrowing Alibaba’s average cost per task.
This also means a token reduction and a cost reduction should remain separate outputs. Your test may find fewer tokens but no economic advantage after prices and retries are applied. Conversely, lower cached-input pricing could produce a cost change larger than the token change. Neither result would contradict the other; they answer different questions.
Quality parity remains an open question
Alibaba reports Work360 scores of 52.69 for Qwen4-Max and 47.74 for Qwen3.8-Max in its latest evaluation. Those are vendor-reported quality results, obtained with the same unpublished methodological details that limit the token claim.
The only independent measurement in the supplied research points in more than one direction. In StealthModels’ 11 November 2026 StealthMark run, Qwen4-Max-Preview scored higher than Qwen3.8-Max in visual intelligence, led by Mona Lisa and Selfie, but lower in general intelligence, where both games fell. That result does not measure coding-task cost or token consumption, and a preview result should not be treated as definitive for every release surface. It does directly weaken any assumption that fewer tokens imply uniform quality parity.
For a coding team, the quality gate should therefore be task-specific. A switch is justified only when Qwen4-Max meets the same acceptance threshold on your fixed tasks and its fully accounted cost is lower. If accepted completions decline, the relevant comparison is no longer the headline token reduction.
Check release status before treating the test as a rollout plan
There is also an unresolved timeline conflict. StealthModels reports an 11 November 2026 preview ending on 25 November, while the fetched Alibaba blog page carries a 3 August 2026 publication date. The supplied material does not resolve how those dates and labels relate to the gateway route.
Confirm the current status and terms of the exact endpoint before relying on the result. Do not describe the combined sources as a clean general-availability story, and do not assume that a preview window ending changes alibaba/qwen4-max on AI Gateway without evidence tying the two together.
I would not switch a production coding agent on the supplied evidence alone. I would use approximately 24% as the vendor-based Xhigh budgeting assumption, run the constant-effort worksheet, and adopt Qwen4-Max only if current confirmed prices, all retries and accepted-task quality preserve the advantage. If failed attempts or longer traces erase it, the correct result is to keep the existing model. What the evidence supports is a better measurement decision, not a verified 35% saving.
Frequently Asked Questions
What is the AI Gateway model ID for Qwen4-Max?
The AI Gateway model ID is alibaba/qwen4-max.
What is the default reasoning effort for Qwen Code v0.26.5?
StealthModels reports that Qwen Code v0.26.5 gives the qwen4 model IDs the same reasoning levels as Qwen3.8-Max: low, medium and xhigh, with xhigh as the default.
What is the vendor-reported cost reduction per Work360 task?
Alibaba reports that average cost per Work360 task fell from $4.43 to $1.67, a vendor-reported 62% reduction.

Join the discussion
Share a useful perspective or ask a question about this article.