Alibaba says the average hard task on its own benchmark cost $4.43 with its last flagship model1. With Qwen4-Max, generally available on QwenCloud since November 16, 2026,2 Alibaba says the average Work360 task costs $1.67. That is a 62% drop3. The model also holds up to 1 million4 tokens of context and reads images natively, according to Alibaba.
The catch is who did the measuring. Alibaba ran the test on a benchmark it built. Deyron Labs found no independent evaluation of the model on launch day. Sanchit Vir Gogia, chief analyst at Greyhound Research, summed it up: "Alibaba has published the mileage it got on its own road."
The price sheet is harder to argue with. Alibaba Cloud lists Qwen4-Max at $1.50 per million input tokens and $4.50 per million output tokens in Singapore (International)5, down from $2 and $6 for Qwen3.8-Max6. Cheap capability changes where builders lose money. Once the model is the cheap part, the bill depends on what you stuff into the window and on whether you can switch providers next quarter.
Two dials set every task bill
Every model bill is two numbers multiplied together: tokens per task and price per token. A token is the chunk of text a model bills by, usually a piece of a word. Alibaba controls the price dial. The builder controls most of the token dial.
| Average Work360 task2Alibaba · vendor benchmark | $1.67 |
|---|---|
| 20-turn coding agent task6Koda estimate · 200,000 tokens resent per turn | $6.00 |
| Context windowAlibaba · text and native images | 1M tokens4 |
| Trimming 100,000 full-window promptsKoda estimate · list input rate | $135,000 |
With this launch, Alibaba moved both. On price, Alibaba Cloud's list rates fell 25% for input and output alike7. GuruFocus reports Alibaba also added large discounts for cached inputs. Caching means the provider keeps a prompt prefix it has already processed, so resending that prefix bills at a lower rate.
On tokens, Alibaba says Qwen4-Max used about 35% fewer tokens than Qwen3.8-Max on its Work360 benchmark at High reasoning effort3. At the heavier Xhigh setting, Alibaba puts the saving at about 24%. On Terminal-Bench 4.0, a coding-agent test, Alibaba reports savings of about 48% and 49% at Xhigh and High reasoning effort8, respectively.
Multiply the two dials and the math almost closes. A 25% price cut times a 35% token cut leaves 0.75 × 0.65, or about 49% of the old cost8. That's a 51% saving, against Alibaba's reported 62%.3 Where the other 11 points come from is unclear. Caching could explain it, or the way Work360 counts cost could, and Alibaba has not broken it out.
Accounting moves the number too. TechNode reported the same comparison in renminbi, RMB 26.58 down to RMB 11.16, which works out to about 58%. Treat a vendor's headline saving as a starting estimate for your own spreadsheet.
Where the million tokens actually cost money
A 1-million-token window is a bigger suitcase on an airline that still charges by the kilo. You can pack everything, and you pay for every kilo. At Alibaba's list price, one full-window prompt costs $1.50 before the model writes a word.
Alibaba has published the mileage it got on its own road.
Send that full window 100,000 times and input charges alone reach $150,000 at Alibaba's list rate. At the old $2 rate, the same job cost $200,000, so Alibaba's price cut saves $50,000. Now trim each prompt to the 100,000 tokens that matter. The bill falls to $15,000. That $135,000 saving is 2.7 times bigger than the one Alibaba handed you.
Agents make it worse, because most agent loops resend their history every turn. Picture a coding agent that runs 20 turns and resends 200,000 tokens each time. That comes to 4 million input tokens, or $6.00 per task at Alibaba's list price, more than three times Alibaba's $1.67 benchmark average3. The benchmark that counts is your workload.
The cheapest token is the one you never send, so step one is ordering the prompt. Put the stable material first (system instructions and tool schemas) so the cache can reuse that prefix on every call. Step two: retrieve only the documents the current step needs, and drop tool output the agent has already read.
Step three: cap the output length. On QwenCloud's International tier, output costs three times as much as input, $4.50 against $1.50.9 Per token, a rambling answer burns money faster than a long prompt does.
Portability is the other half of the job. OpenRouter lists Qwen4-Max at about $1.425 per million input tokens and $4.275 per million output9, roughly 5% under the $1.50 and $4.50 standard prices in its announcement5. Same model, different door, different price. Only a team that can swap doors collects the difference.
An OpenAI-compatible endpoint works like a travel plug adapter. The prongs fit the socket, but nothing promises the voltage matches. Tool calls parse differently between providers, and structured JSON output breaks in new places.
Closed weights tighten the lock. The Next Web reported that Qwen4-Max ships without open weights, so a team cannot self-host it if QwenCloud changes prices or quotas. TechNode reports Alibaba plans to open-source smaller Qwen4 models, which may give a partial fallback later.
We think the adapter layer is the most underbuilt piece of most AI stacks. It needs one internal request format with a translator per provider, plus a test set of real prompts that runs against each translator.
Sub-dollar tasks may arrive next generation
The trend line from this launch is short, and it is steep. In one generation, Alibaba says, Work360 task cost fell from $4.43 to $1.67.3 Alibaba kept the 2.4-trillion-parameter10 backbone and cut active parameters per token from about 95 billion11 to about 71 billion. Fewer active parameters means less compute behind each token.
If the next Qwen generation repeats a 62% cut, the same task lands near $0.63. One more repeat puts it near $0.24. These are napkin projections from one vendor's one benchmark, and nobody should budget on them. The direction still matters, because the price gap across vendors is already wide.
Infosec Today reports Anthropic lists Claude Opus 5.5 at $20 per million output tokens12, about 4.4 times Qwen4-Max's $4.50. A gap that wide is leverage in every contract renewal, but only for buyers who can move a workload in a week.
Capability still matters. Alibaba reports the share of tasks passing both function and security checks on Last Secure Code Bench rose from 22.0% to 31.8%.10 So about two in three tasks still fail those checks.
Gogia's point is that the real unit is the cost of an accepted result, counting retries and the staff time spent checking outputs. A cheap model that needs a human to fix every third answer can cost more than a pricey one that ships clean. The evidence points to a clear pattern: over the next few model generations, the winners will be teams that measure accepted results per dollar on their own tasks.
The asymmetry favors building the exit now. An adapter and a 50-prompt test set cost a few engineering days, once. Without them, a team keeps paying its current provider's rate every time prices fall somewhere else.
Price one real task on two providers
Step one: pick one task your product runs every day, such as a support reply or a code review. Run it 50 times and log input tokens and output tokens for each run. OpenAI's APIs return both counts in the response.
Step two: price each run at Qwen4-Max's list rates and at your current provider's list rates. At Alibaba's prices, a run with 30,000 input tokens and 2,000 output tokens costs $0.045 plus $0.009, or $0.054. Fifty of those runs cost $2.70.
Step three: mark which outputs you would actually ship, then divide total spend by accepted runs. If 40 of the 50 pass, the cost per accepted result is $2.70 divided by 40, or about $0.068. That number, per provider, is the one to compare.
Step four: reorder the prompt so the stable instructions and schemas sit at the top, untouched between calls. Check Alibaba Cloud's pricing page for the current cached-input rate. Then rerun the 50 and see how much of each prompt bills as cached.
Step five: wrap both providers behind one function that takes your internal request and returns text plus token counts. Run the same 50 prompts through it every week. Expect breakage on the first pass, such as tool calls that fail to parse or JSON with a missing field. Save every failure to a file, because that file is your migration map when prices move again.
Practise it step by step: Build the Verification Habit
The bottom line
Cheap tokens pay off only for teams that measure and move.
Qwen4-Max lowers the price dial, and Alibaba's own numbers still need checking on your workload. The larger savings come from sending fewer tokens and capping output. A few engineering days spent on an adapter and a shared prompt test set let a team follow prices as they fall. We would build that exit now, while switching is still cheap.
