Qwen4-Max sells 62%-cheaper tasks on its own benchmark; the independent index reads 11%
Alibaba claims Qwen4-Max makes an agent task 62% cheaper. The independent index measures 11% against Kimi K3.1 Pro at 31.9 minutes, and no weights ship.

Alibaba published Qwen4-Max on 15 November, and the launch post leads with cost. An agent task on its Work360 benchmark now averages $1.67, against $4.43 for Qwen3.8-Max, a fall of roughly 62%. That figure has already hardened into a description of the model. It is a description of one benchmark, run by the company that sells the API, on a harness the company built. The independent profile at Artificial Analysis puts the same model about 11% below Kimi K3.1 Pro on cost per task, exactly at the median on intelligence, and notably slow per task. Both numbers can be honest. They measure different questions.
A 2.4-trillion-parameter model that computes 71 billion of them
The backbone is 2.4 trillion parameters with 71 billion activated per token, plus 328 billion N-gram embedding parameters. The distinction matters. Those embeddings are lookup tables, read from memory rather than multiplied through the forward pass, so counting them as parameters is accurate without inflating the arithmetic bill. The 71 billion is the number that decides speed and price.
The architecture is unusual. Gated DeltaNet layers alternate with Qwen Sparse Attention, whose indexer selects at micro-block granularity rather than token by token, so the attention budget goes where the signal is likely to be. A Gated Residual with four parallel branches replaces the plain residual stream, and an N-gram Embedding module sits at layer two, injecting memorised continuations before the network has done much work. Vision is native rather than bolted on, context reaches one million tokens, and Alibaba reports no loss spikes during training.
The base checkpoint is where the efficiency story is cleanest: lowest bits-per-byte of five checkpoints evaluated, 5 to 13% below Qwen3.8-Max-Base while activating 25% fewer parameters. That is a pretraining claim, measured on held-out text, and it does not depend on an agent harness.
Where 62% actually comes from
Work360 is Alibaba's own agentic benchmark: 26 tier-one domains, 182 subdomains, 721 specialised fields, with tasks drawn from professional workflows rather than exam questions. Qwen4-Max scores 52.69 against 47.74 for Qwen3.8-Max, a gain of 4.95 points or about 10.4% relative. It is a vendor benchmark, run by the vendor, and nobody outside Alibaba has scored the model on it.
The cost reduction is a product of two inputs. Token consumption falls about 24% at the Xhigh reasoning setting and 35% at High on Work360, and about 48% and 49% on Terminal-Bench 4.0. Price per token is lower than the model it replaces. Multiply the two and $4.43 becomes $1.67. Neither half is wrong; the headline is simply what happens when a vendor multiplies its own two savings and quotes the product.
What the independent harness reads
Artificial Analysis classifies Qwen4-Max as proprietary, released in November 2026, and scores it 43 on its Intelligence Index against 34 for Qwen3.8-Max on the launch chart. That places it 47th of 91 models, with the median of comparable models also at 43. A 2.4-trillion-parameter flagship landing mid-field on the one index that reruns everything itself is the uncomfortable fact at the centre of this launch.
Speed is worse. Time to first answer token is 147.96 seconds, output runs at 55 tokens per second, and a full Intelligence Index task takes 31.9 minutes against a median of 11.3. The run generated 127 million tokens on the index, against a median of 65.4 million. Artificial Analysis's summary is blunt: "below average in intelligence, but well priced when comparing to other models of similar price. It's also notably slow per task and very verbose."
Verbosity and price interact. A model that emits nearly twice the tokens of its peers only partly recovers that through a lower rate, which is how the vendor's 62% becomes $3.27 per index task against $3.68 for Kimi K3.1 Pro, roughly 11%. VentureBeat's analysis makes the same point plainly: the 62% comes from Work360, the halved token count comes from a terminal benchmark, and neither establishes a savings rate for anyone else's workload.
The price sheet
| Model | Input per 1M tokens | Output per 1M tokens |
|---|---|---|
| Qwen4-Max | $1.50 | $4.50 |
| Kimi K3.1 Pro | $2.00 | $10.00 |
| Gemini 4 Hydrogen | $1.50 | $7.50 |
| GPT-6.1 Sol | $2.00 | $10.00 |
| Claude Sonnet 5.5 | $2.00 | $10.00 |
| Claude Opus 5.5 | $4.00 | $20.00 |
The list price is genuinely competitive, and caching strengthens it: cached input runs $0.075 per million, a 95% discount. Cache writes are billed separately at $1.875 for five minutes and $3 for one hour, so a workload with poor cache reuse pays closer to list than the headline discount suggests. Gemini 4 Carbon, absent from the table, is $2 for input and $10 for output as an introductory rate, moving to $4 and $20 on 1 December.
The flagship ships without weights
Qwen built its reputation on downloadable weights. Qwen4-Max has none. Access runs through the QwenCloud API, Qwen Studio, QwenWork and Qoder, and Artificial Analysis lists the model as proprietary. The launch post promises open releases of smaller Qwen4 models but gives no dates, no licences, and no equivalent commitment for Max, a gap already visible when Qwen's open-source pledge went up two weeks ago. Anyone who chose this family because they could run it on their own hardware is now looking at a rented endpoint at the top of the range.
The vendor chart, read carefully
Qwen's comparison numbers are its own and were not independently rerun. DeepSWE v1.1 comes in at 75.7 against 69.3 for the previous model, Terminal-Bench 4.0 at 48.5 against 38.9, APEX-Agents v1.1 at 72.1 against 63.3, AutomationBench-AA at 72.5 against 56.2, ChartoGraphy at 54.2 against 42.0, and Agents' Last Exam v1.1 at 46.7 against 27.9. On that last one Claude Opus 5.5 scores 43.4, which makes it the single strongest line in the post.
ProgramBench is the line to read twice. Qwen4-Max scores 31.5 against 28.0 for its predecessor, and the same table puts Claude Opus 5.5 at 65.0, GPT-6.1 Sol at 52.5 and Kimi K3.1 at 25.0. Against frontier rivals the gap is not four points, it is more than thirty.
What the numbers support
None of this makes Qwen4-Max a failure. A model at the median of an independent index, priced under most rivals, with native vision and a million-token window, is a defensible buy for batch and long-running agent work where latency is not the constraint. The 147-second first token is close to disqualifying for chat and nearly irrelevant to an overnight pipeline.
What the 62% cannot carry is a budget. On Alibaba's benchmark, with Alibaba's harness and Alibaba's pricing, the saving is real and the token reduction is documented. On a suite that counts tokens emitted as carefully as tokens charged, the edge over the nearest well-known rival is about a tenth. Run the workload, count the tokens, then quote the number.


