Qwen4-Max Preview vs Claude Opus 5.5: 4 Tests, Costs, and Access

By Beau Johnson··8 min read

Quick answer

Qwen4-Max Preview kept up with Claude Opus 5.5 on every test it finished, including 12 of 12 hidden tests on a client job. The problem is access: Token Plan only, no pay-as-you-go API, three runs in the Qwen app ended with no answer, and the preview is retired after November 25.

In this video I run Alibaba's Qwen4-Max Preview against Claude Opus 5.5 on my four-test bench on day one. If you build with Claude Code and want to know whether Qwen 4 is worth a look, I cover what it is, how you can actually get it, what each test cost, and my verdict.

What I cover in this video

  • 0:46 What Qwen4-Max Preview is, its settings, and its expiry date
  • 2:20 How you can use it: the Token Plan, the fine print, and no pay-as-you-go API
  • 4:36 What people built on day one, and why I'm careful with those posts
  • 7:39 Test 1: a first-person haunted house game
  • 9:11 Test 2: a real client job graded by 12 hidden tests
  • 10:45 Test 3: the viral pelican on a bicycle, redone with our own prompt
  • 12:17 Test 4: a bug hunt with 16 planted bugs
  • 13:48 The verdict and who should use it

What Qwen4-Max Preview is

Qwen4-Max Preview is Alibaba's next flagship. Their docs describe it as a multimodal model built for coding, office work, front-end design, and complex agent tasks. It takes text, images, and video, has a 1 million token context window, and thinking is on by default. It has three effort settings: low, medium, and extra high. There is no plain high, and Opus runs on high in my bench, so the settings can't line up perfectly.

There is no benchmark table, no model card, no parameter count, and no open weights, so you can't download it and run it locally. It also has an expiry date. Alibaba's docs say the preview runs until November 25, and then "The model will be retired after the preview ends." They don't say what replaces it or when the full Qwen 4 Max ships.

There is no independent score either. Artificial Analysis has no Qwen 4 row yet. The older Qwen3.8 Max scores 34 on their Intelligence Index, while Opus 5.5 scores 52 on high effort and 59 on max. So Qwen 4 has a big gap to close.

The biggest day one posts look great, but a lot of them end with nearly the same sentence about a limited time discount in Qoder, Alibaba's coding agent. I can't tell you anyone was paid. The cleanest comparison I found was one creator's Crossy Road game, where their verdict was "I think Opus won this one, but it was definitely close."

How you can use it, and how I tested it

Access is a mess. Alibaba's docs say it is Token Plan only, with no pay-as-you-go access, and it isn't on OpenRouter. The Token Plan is a monthly subscription:

  • Lite $6, Essential $10, Standard $18, Pro $43
  • New users can get Lite for $3 until November 30
  • During the preview, Qwen 4 burns credits at as little as 20% of the normal rate

The fine print matters. The plan is for interactive use inside tools like Claude Code, Cursor, and Qwen Code. Alibaba says it must not be used for automation scripts or batch jobs, and breaking that can get your key banned. Right now you also can't cancel the personal plan once you're in.

I paid for the Qwen app's Plus plan at $19 a month and checked whether Qwen Code would let me run the preview in the terminal. It doesn't. On Plus, Qwen Code only offers the older Qwen3.8 Max and Flash. So Opus 5.5 ran as a coding agent in Claude Code on high effort with web search off, and Qwen ran in the Qwen app's chat in Thinking mode with no terminal. It can't run or test what it writes, so build times are not like for like.

On costs, Opus shows its API-equivalent price. Qwen 4 has no published price, so I priced what it visibly wrote at Qwen3.8 Max's rate of $2 in and $6 out per million tokens. Hidden thinking isn't counted, so the real number would be higher.

The four tests

Haunted house game

Opus took 35.9 minutes and $4.72, including one restart from our 15 minute check-in. Qwen took 22.1 minutes for an estimated $0.31. Partway through, Qwen wrote "Wait, I made a structural error mid-file. Here is the corrected complete file." and then wrote the whole game again in the same answer. It needed two tries, because the app's Web Dev mode ended at 27 minutes with no file. I think Qwen's hallway is creepier right out of the gate.

A real client job

This is a messy email thread from a design studio owner who wants invoices and payment reminders, and who changes the tax rule in the last email. Both models scored 12 of 12 on hidden tests, both caught the tax change, and both left the vague late fee out and asked her to confirm. Two judge models, Kimi K3.1 Pro and Grok 4.7, gave both reply emails 10 out of 10. Opus finished in 8.3 minutes for $2.04. Qwen took 30.6 minutes for an estimated $0.15.

The pelican on a bicycle

The most viewed Qwen 4 test we found on X was a 3D pelican whose bike wheels flip like spinning coins. The author never shared the prompt, so I wrote my own. Qwen's first try ended at 18 minutes with no answer. The second finished in 23.6 minutes for an estimated $0.18, and the coin-flip wheels didn't show up. Opus took 22.2 minutes and $5.40. Opus' scene has more life in it, and Qwen's has more controls.

The bug hunt

This is an invoicing backend, about 3,000 lines across 42 files, with 16 planted bugs and 11 decoys. Opus fixed 15 of 16 with zero regressions in 8.3 minutes for $2.24. Qwen has no result. We pasted the whole codebase into the chat twice, and both times the app put a slide puzzle check in front of it and the request never started. I don't count that as Qwen losing, because the app never took the codebase.

My verdict

The model kept up every time it got to answer. Getting it to answer was the hard part. Three of our runs in the app ended with no answer, you can't buy it by the token, and the plan you can buy bans automation. For my resale business, the work I'd hand a model runs on a schedule, like pulling numbers after every stream, so I can't use it yet. Opus 5.5 in Claude Code stays my daily driver. If you want more options for building, I keep a list of the best vibe coding tools.

Key takeaways

  • On the one test with a hidden answer key that it got to answer, Qwen4-Max Preview tied Opus 5.5 at 12 of 12 with a perfect reply email.
  • Access is the real problem: Token Plan only, no pay-as-you-go API, and no automation scripts or batch jobs allowed.
  • The preview is retired after November 25, so don't build anything on it that needs to exist in December.
  • If you're already in Qoder or on the Token Plan and want a strong second model for front-end work at a discount, try it while the preview is live.
  • These are one run each from November 11, with Qwen in a chat app and Opus as a coding agent, so treat it as a vibe check.

FAQ

Is Qwen4-Max Preview as good as Claude Opus 5.5?

On every test it finished in my bench, it kept up. Both models scored 12 of 12 on hidden tests for a real client job and got 10 out of 10 on the reply email from both judge models. Qwen has no result on the bug hunt because the Qwen app never started the request.

How do you get access to Qwen4-Max Preview?

Alibaba's docs say it is Token Plan only, with no pay-as-you-go API, and it isn't on OpenRouter. There is also early access in Qoder and in the Qwen app, which is where I ran it on the $19 a month Plus plan.

How much does Qwen4-Max Preview cost?

Alibaba hasn't published a price per token. The Token Plan costs $6 for Lite, $10 for Essential, $18 for Standard, and $43 for Pro per month, and new users can get Lite for $3 until November 30.

When does Qwen4-Max Preview end?

Alibaba's docs say the preview runs until November 25, 2026, and the model is retired after the preview ends. They don't say what replaces it or when the full Qwen 4 Max ships.

Chapters

  1. 0:00Qwen4-Max Preview vs Opus 5.5: it kept up, then the app got in the way
  2. 0:34What Qwen4-Max Preview is (and its expiry date)
  3. 2:13Beckon: every coding agent on one canvas
  4. 2:28How you can actually use it (Token Plan, no API)
  5. 3:44The independent numbers (there aren't any yet)
  6. 4:51What people built on day one
  7. 6:32How we tested: a coding agent vs a chat app
  8. 7:53Test 1: the haunted house game
  9. 9:39Test 2: a real client job (hidden tests)
  10. 11:06Test 3: the viral pelican on a bicycle, redone
  11. 12:23Test 4: the bug hunt (16 hidden bugs)
  12. 13:57The verdict

Transcript

Read the full transcript

0:00Alibaba just opened Quen 4 Max preview and on every test it finished, it kept up with Claude Opus 5.5. 12 out of 12 on a hidden client job. And the viral pelican that broke physics rode just fine in our run. But quality wasn't the problem. Three of our runs never came back and you can't buy it the normal way. In this video, I'm covering what it is, the twoe catch, the day one builds, and four tests with what each one costs. Opus builds every video on this channel, so I had to know. Let's jump right into it. So, what is Quen Formax Preview exactly? It's Alibaba's next flagship. Their docs call it the next generation flagship multimodal model in the Quen family. Built for coding, office work,

0:46front-end design, and complex agent tasks. It takes text, images, and video. It has a 1 million token context window, and thinking is on by default. And it has three effort settings, low, medium, and extra high. There's no plain high, which matters later because Opus runs on high in our bench. So, right away, the settings can't line up perfectly. And the launch post just said, "Quen 4 is coming with three promises: deeper knowledge, richer world understanding, and stronger coding and agent skills." It got over a million views, and there's not one number in it. The last line just says, "Put it to the test on a real problem. Show us what you build." So that's exactly what we did. There's no benchmark table, no model card, no

1:33parameter count, and no open weights. So if you were hoping to download this one and run it locally, you can't. And this one surprised me. This model has an expiry date. Alibaba's own docs say the preview runs until November 25th, and then quote, "The model will be retired after the preview ends." That's about 2 weeks. They don't say what replaces it or when the full Quen 4 Max ships. And to be clear, only the preview is out. Alibaba's own post says Quen 4 is coming. So, you're looking at a flagship with no numbers, no price per token, and a 2 week shelf life, which is exactly why I wanted to test it myself. Still juggling cloud code, codecs, and cursor in five terminal windows. Beckin puts every coding agent on one canvas side by

2:20side. Press a key and just talk to them. Every idea gets its own workspace with memory that carries over. Becken, hey, beckon.ai. So, how do you actually get it? Honestly, it's a mess. There's no pay as you go API. Alibaba's docs say it straight. Token plan only. Pay as you go access is not supported. It's not on open router either. The token plan is a monthly subscription. Light is $6, Essential 10, Standard 18, and Pro 43. New users can get Light for $3 until November 30th. And during the preview, Quen 4 burns your credits at as little as 20% of the normal rate. Sounds cheap, right? Now, the fine print. That plan is for interactive use only inside tools

3:06like Claude Code, Cursor, and Quen Code. Alibaba says it must not be used for automation scripts or batch jobs. And breaking that can get your key banned. And right now you can't cancel the personal plan once you're in. Credits run out and you wait until next month. There's also early access in coder Alibaba's own coding agent and in the Quen app. So I paid for the Quen app's plus plan, $19 a month, and checked if Quen Code would let me run it in the terminal. It doesn't. On plus, Quenode only offers the older Quen 3.8 ate max and flash. So, that decided how we tested it, and you'll see why that matters when we get to the tests. So, what do the referees say? This is usually where I pull up artificial analysis and show you the score. There

3:51isn't one. I checked again right before writing this. No Quinn for row on the leaderboard and the model page doesn't exist yet. No independent score. What we do have is where Alibaba was before this. The old flagship Quen 3.8 8 max scores 34 on the artificial analysis intelligence index. Opus 5.5 on high effort which is what we run in our bench scores 52 on max 59. So Quen 4 has a big gap to close. And Quen 3.8 Max's tokens cost about a third of Opus on output. $2 in six out versus 4 and 20. But to run one task on that index it cost $102. Opus on max $104.

4:36The same cheap tokens didn't make a cheap task. Keep that in mind because Quen 4 has no price at all yet. So when I show you Quen's cost in our tests, I'm using that old Quen 3.8 max price as a standin and I'll label it every time. So what did people actually build with it on day one? The biggest posts look amazing. a storywriting app, a 3D downhill game with a fox, a car showroom, a restaurant website. But scroll through them for a minute and you'll notice something. A lot of them end with almost the exact same sentence. Now available in coder, Alibaba's coding agent with a limited time 80% discount. I can't tell you anyone was paid, but a lot of the first day posts read the same, and some of them aren't even this

5:22model. Before launch, COD ran a mystery model for 4 days called Koduck, an early Quen test build. Some of the clips posted as Quen 4 still say Koduck right on the screen. And that Fox game was already in Coder's own ad 2 days before launch. The cleanest comparison I found was one creator's Crossy Road game, Quen 4 versus Opus 5.5 side by side. Their verdict, quote, I think Opus won this one, but it was definitely close. And to be fair, they ran Opus on Medium. And then there's the doubt. One tester ran the same prompt in two different tools and said they honestly hoped it was wired to the wrong endpoint, but the API answered as Quen 4, so they doubted it. And the most viewed test of Quen 4 we found on X at about 195,000 views when

6:09we checked was a fail. A 3D pelican riding a bicycle where the wheels flip around like spinning coins. The author's caption translated says, "Something is definitely off." People in the replies asked for the prompt. It was never shared. Somebody else got a pelican walking with the bike strapped to its back. So, was that the model or one bad run? That's our third test. So, here's how our bench works this time, and it's different from usual. Normally, every model runs as a coding agent in its own folder where it can write files, run them, and fix them. Opus 5.5 ran exactly like that in cloud code on my subscription on high effort. But Quen Formax preview isn't available to coding agents on the plan we have or through a

6:54pay as you go API. So it ran in the Quen apps chat on the $19 plus plan in thinking mode. No terminal. It can't run or test what it writes. It answers once and that's it. We pasted the same prompts into the chat plus one line asking for the full file in one code block. So build times aren't like for like a chat answer versus an agent that runs and checks its own work. And the app had web search switched on. Opus had it off. Costs. Opus shows its API equivalent price, what those tokens would cost on Anthropic's list price. Quen 4 has no published price. So I'm counting what it visibly wrote and pricing it at Quen 3.8 Max's rate. Its hidden thinking isn't counted. So the real number would be higher. Four tests.

7:39A haunted house game. A real client job with hidden tests, that viral pelican, and an extreme bug hunt. One run each and I'll tell you every time Quinn needed a second try. This is a vibe check, not a verdict from a lab. Test one, a firsterson haunted house horror game. A flashlight, three glowing keys, a ghost that drains your sanity, and it's supposed to be creepy in the first 10 seconds. Opus first 36 minutes, and it took a while to get going. It hadn't saved anything after 15 minutes, so our harness restarted it once in the same folder. That's a rule we use for every agent run. But look at what it built. A moonlit hall, a chandelier, a flashlight cone, a sanity bar, and a ghost. The

8:24hint says, "Keep your light on her and she'll retreat." And that portrait close-up is creepy. A PI equivalent cost $4.72. Now, Quen, 22 minutes. Its game opens on a title screen. So, the recorder pressed the start button first, then the same inputs as Opus, and within seconds, you can see a ghost standing at the end of the hallway. Grandfather clock, staircase, chandelier. Zero errors on the page. And this part is wild. After it finished a whole first file, it wrote, "Wait, I made a structural error midfile. Here is the corrected complete file." And then it wrote the whole game again in the same answer, 99 kilob. We used the corrected one. Estimated cost about 31 cents at Quen 3.8 Max's price

9:11and that's only what we could see, but it took two tries. The first time I used the app's webdev mode. 27 minutes later it ended with the request is ended and no file. The plane chat run is the one that finished. That's the app, not the model's answer, but it's what you'd run into. So, which one is scarier to you? I honestly think Quen's hallway is creepier right out of the gate. Comment down below. And that's the fun test. The next one is the one that would actually pay my bills. Test two is a real freelance job. A messy email thread from Dana, a onewoman design studio who wants invoices and payment reminders. She asks for 8 and 1/4% tax on every invoice. Then in the last email, she gets off the phone with her accountant. Design work isn't taxable. Only the printing lines

9:58get tax. And there's one vague bit, a late fee of 1.5%, but she's not sure if it's monthly or one time. The model has to build the app and write the email back to Dana. Then 12 hidden tests the models never saw. Check everything. Invoice numbers, due dates, reminders, payments in chunks, the rounding, and that tax change. Opus 12 out of 12. Quen 12 out of 12. Both caught the tax change. Both handled the late fee the same smart way. They left it out and asked Dana to confirm. The reply emails were scored by two judge models that aren't in this matchup. Kimmy K 3.1 Pro and Grock 4.7 against a hidden rubric. Opus 10 out of 10 from both. Quen 10 out of 10 from both. A dead tie. The

10:45difference is the clock. Opus finished in about 8 minutes for $24 API equivalent. Quen took 30 and 12 minutes in the app on the first try for about 15 cents at the stand-in price. So on the work I actually get paid for, the preview kept up completely, which made me even more curious about that pelican. Test three, the pelican on a bicycle. The original author never shared their prompt, so this one's ours. A pelican riding a bicycle along a coastal road at sunset in 3D with follow, side, front, and topown cameras. And the app got us again. Quen's first try ran 18 minutes and ended with the request is ended and no answer. The second try, same prompt,

11:31finished in about 24 minutes. So here's Quen's. A lighthouse, the sun setting on the sea, a headlamp beam on the road, and a big control panel. Look at the side view. the pelicans on the pedals and the bike looks right. The coin flip from the viral clip didn't show up in any of the frames we looked at. Estimated cost about $18 at Quen 3.8 Max's price, counting only what it wrote. And Opus, 22 minutes, $5.40 API equivalent. The Pelican has a little basket. There are seagulls, rocks out in the water, and the bike is right from every camera. One button cycles through all four views. So, was the viral one the model or one bad run? From what we saw, one bad run or a prompt we'll never

12:17see? Opus' scene has more life in it. Quen's has more controls. Which one would you rather ride? And now the hard one. Test four, the bug hunt. This is a real code base. An invoicing backend for a small business. About 3,000 lines across 42 files. We planted 16 bugs in it. Most of them start in one file and only show up in another. A double charge when two payments land at the same time. Tax rounded wrong. A security hole where an API key from one company can read another company's invoices. Plus 11 decoys that look like bugs but aren't. And the models weren't told how many bugs there are. Hidden tests decide what's actually fixed. Breaking something that worked costs points. Opus

13:02took a little over 8 minutes for $2.24 API equivalent. It fixed 15 of the 16 bugs with zero regressions. It even fixed the cash race, a bug where a cash gets filled at the same moment it's being cleared. Opus missed that one on our last bench. The only one it missed this time is a refund that should turn payment reminders back on. And Quen, this one has no Quen score. A chat app can't open a project folder, so we pasted the whole codebase into the chat. 42 files, about 128,000 characters. We sent it twice and both times the app put a slide to complete the puzzle check in front of it and the request never started. We don't solve those puzzles in our bench. So this test just doesn't have a Quen result. That's not Quen losing the bug hunt. The app wouldn't

13:48take a full codebase from us. And honestly, that sums up this whole bench. The model kept up every time it got to answer. Getting it to answer was the hard part. So who wins on the one test with a hidden answer key that it got to answer? Quen for max preview tide opus 12 out of 12 and a perfect reply email. Its haunted house was creepier and it's Pelican Road just fine. But the story of this video is access. You can't buy it by the token. The plan you can buy bans automation. It isn't in Quen code on the plan I paid for. It ran in a chat app with no terminal and three of our runs in that app ended with no answer after 18 to 45 minutes. And the app wouldn't even take our bug hunt. And in two weeks, this exact model gets retired.

14:34So, who should use it? If you're encoder or on the token plan already and you want a strong second model for front-end work at a discount, try it while the preview's live. Just don't build anything on it that needs to exist in December. For my resale business, I can't use it yet. We sell live on whatnot and Tik Tok shop, and the stuff I'd want to hand a model runs on a schedule, like pulling numbers after every stream. That's exactly the kind of automation the token plan bans. So, Opus 5.5 in Cloud Code stays my daily driver. Still, if the full Quen 4 Max ships with a real API and a real price and it's this good, Anthropic has something to worry about. So, I'm curious, would you build on a model that's only around for 2 weeks or wait for the real release? Let me know down below. And if this

15:19helped you figure out where Quen 4 fits, a like really helps the video spread. Have a blessed day.

Ready to start building with AI?

Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.

Join Shipping Skool

Keep reading