Copilot Pay-As-You-Go Pricing: Ignite 2026 Verdict
Alibaba opened a preview of Qwen4-Max on November 11, 2026, with a 1M-token context window, a 393,216-token output cap, and a hard stop at the end of November 25. You can’t call it with an API key and a credit card. You can reach it only through Alibaba’s Token Plan subscription.
Two weeks, one subscription, no benchmark table. That’s the offer. So is it worth paying to look?
Short answer: for most teams, probably not, and the reason is in Alibaba’s own documentation. More on that below.
Quick Summary: What Happened
Detail Info Model ID qwen4-max-previewWindow 10:00 Nov 11 to 24:00 Nov 25, 2026 (UTC+8), per Alibaba’s docs Access Token Plan only. No pay-as-you-go. Context / output 1M tokens in; 393,216 tokens out, thinking included Input Text, images, video Not published Benchmarks, model card, parameter count, license, weights Official pricing No per-token price on the preview page Bottom line: A real model with real specs and zero public evidence of quality. Run a trial only if you already have a test suite and a spare afternoon.
Alibaba’s own preview page is the only source I’d treat as settled. It says the following.
qwen4-max-preview. The same page says the model retires when the preview ends.reasoning_effort set to low, medium or xhigh (xhigh is the default).That last point matters more than it looks. Anthropic-format compatibility means you can aim an existing agent harness at it without a rewrite. Low switching cost is the one genuinely good thing about the setup.
Here’s what Alibaba hasn’t published: no benchmark table, no model card, no parameter count, no license, no weights. AI/TLDR’s release note says the Qwen team claims stronger knowledge, coding and agentic skills than its predecessor Qwen3.8, and that no outside evaluator has checked it.
“Better than our last model” is not a number. It’s a sentence.
Some other claims floating around:
Two details change the math.
First, Alibaba’s preview page says the production version, Qwen4-Max, is available and recommended. AI/TLDR links a separate release headline claiming Qwen4-Max is live on the API at $1.50 input and $4.50 output, but that price is unverified, and the aggregator doesn’t verify it either. If a production model exists, a preview that dies on November 25 is a poor place to build anything.
Second, per Xenospectrum’s reading of the docs, Token Plan is for interactive use. It reportedly prohibits automation scripts, in-house backends and non-interactive batch processing, and is sold only in the Singapore region. Credits drain with token count, thinking and tool calls. And the docs say global-environment data may cross borders during inference. If your data policy cares about that, stop here.
So a trial means a person, in a coding tool, poking at it. Not a pipeline. Not an eval harness running overnight (read the terms before you try).
Here’s how I’d decide.
Fewer than three yeses? Wait.
There is nothing to compare against. No published coding score, no agent benchmark, no independent run. Anyone writing “Qwen4 beats Claude” this week is guessing. Our guide to the best AI coding assistants shows what a real comparison needs: the same tasks, same harness, scored by someone other than the vendor.
If you do trial it, run Claude on the identical ten tasks. Otherwise the result tells you nothing.
Unknown. Alibaba’s preview page lists no weights and no license. Some earlier Qwen models did ship open, and we covered the Qwen 3.5 open-weight release. But the Max line has been a hosted flagship, and nothing published says that changes. Treat open-weight Qwen4 as a rumor until Alibaba says it.
The roadmap is in the same state. Naming and the “full Qwen4 series” talk come mostly from secondary outlets. Xenospectrum lists the full series release as unpublished. Plan around what’s in the docs.
I think the preview is a smart marketing move and a weak buying signal. Alibaba gets developers into Token Plan, a subscription product, without the scrutiny a public benchmark table brings. You get a closing clock and a discount. Urgency is a sales tool, not an evaluation method.
Is the model good? Possibly. Qwen has earned some benefit of the doubt, and our DeepSeek V4 review shows Chinese labs like DeepSeek can ship strong, cheap models. But “possibly good” doesn’t justify rearranging a roadmap. If cost matters, our cost optimization guide is a better starting point than a 14-day subscription.
One more thing. A retired model ID means any experiment you build breaks on November 26. Whatever you learn has to transfer to the production model, and you can’t know that it will.
It’s Alibaba’s preview of its next Max-tier model, available since November 11, 2026 under the ID qwen4-max-preview. It has a 1M-token context window and accepts text, image and video input, per Alibaba’s documentation.
At 24:00 on November 25, 2026, UTC+8, according to Alibaba’s docs. One outlet said November 24, but the official page doesn’t.
There’s no per-token price on the preview page. Access requires a Token Plan subscription, and reported discount figures conflict between sources. Check Alibaba’s Token Plan overview before buying.
Not for the preview. Alibaba says pay-as-you-go isn’t supported. The page does point to a production Qwen4-Max, but its pricing is unverified.
No published ones. Alibaba released no benchmark table, model card or parameter count, and the Qwen team’s claims of gains over Qwen3.8 haven’t been independently checked.
Nobody has said. No weights or license exist for the preview, and the open-weight question is speculation for now.
Last updated: November 16, 2026. Sources: Alibaba Cloud — Qwen4-Max-Preview · Alibaba Cloud — Token Plan overview · Xenospectrum — Qwen4-Max-Preview specs and limitations · AI/TLDR — Qwen4-Max-Preview · FourWeekMBA — Alibaba opens a 350-hour Qwen4-Max preview.
Related reading: Best AI Coding Assistants 2026 · Qwen 3.5 Review · DeepSeek V4 Review · AI Cost Optimization Guide