Deep Dive — Qwen4-Max vs Opus 5.5: Alibaba is 16 points short

Share
Deep Dive — Qwen4-Max vs Opus 5.5: Alibaba is 16 points short

Alibaba released Qwen4-Max today, the first model built on its new Qwen4 architecture: a 2.4-trillion-parameter mixture-of-experts backbone that activates 71 billion parameters per token, plus a separate 328-billion-parameter N-gram embedding table, native vision, and a one-million-token context window, according to the Qwen team's launch post. The API price is $1.50 per million input tokens and $4.50 per million output, with cached input at a 95% discount.

The number that will decide whether you switch is not any of those. It is 43.

That is Qwen4-Max's score on the Artificial Analysis Intelligence Index, per its model page — nine points above its own predecessor, Qwen3.8-Max, and 16 points below Claude Opus 5.5 at 59. It also lands three points under Kimi K3.1 Pro, a model whose weights you can download. Alibaba's biggest model ever is not the best model in the world, and it is not even the best Chinese model once you weigh what the open-weight camp ships.

To be clear about who should care: if you are running long agent jobs where a percentage point of task success is worth real money, Qwen4-Max is a plausible fourth option behind Opus 5.5, Gemini 4 Carbon and Gemini 4 Hydrogen. If you are Alibaba, this is the model that has to justify a decade of Qwen brand equity built almost entirely on giving weights away.

What actually shipped

The architecture is the most interesting part, because it is where the token savings come from. Three of every four attention layers are Gated DeltaNet, which compresses conversation history into a fixed-size state instead of keeping every key and value around. The fourth is Qwen Sparse Attention, which uses a small indexer to pick relevant blocks of context rather than attending to everything. The residual stream is split into four parallel branches with learned read and write gates, and a lookup table at layer two adds representations of common token patterns for almost no extra compute.

The payoff shows up in pre-training and in token counts. Alibaba says the 2.4T-parameter run had no loss spikes, and that Qwen4-Max-Base beats its predecessor on bits-per-byte across knowledge, code, question answering and STEM by 5% to 13% while activating 25% fewer parameters per token. On its Terminal-Bench 4.0 coding evaluation, the same score takes 151K tokens per attempt instead of 289K — a 48% cut — while the score rises from 38.9 to 48.5.

That is a real engineering result, and it is the honest core of the launch: Alibaba made a model that thinks in fewer tokens, which is the only kind of efficiency that survives contact with an invoice. Everything else in the announcement is downstream of it.

The benchmark sheet, read the way Alibaba would rather you didn't

Alibaba published sixteen-odd evaluations, and Qwen4-Max wins a lot of them. It scores 75.7 on DeepSWE v1.1, 85.0 on BioMysteryBench, 54.2 on Chartography (up 12.2), 68.5 on JobBench, 21.3 on the legal research benchmark HLAB-Vals-1.1, and 52.69 on Work360, Alibaba's own 721-subdomain model of professional work. On Code Arena's blind front-end leaderboard it sits sixth, between GPT-6 Astra and Claude Fable 5.1 — the best placing a Chinese model has had there.

Now read the column that isn't in the headline. On Terminal-Bench 4.0, its 48.5 trails Claude Opus 5.5, GPT-6.1 Sol and Kimi K3.1 Pro. It scores 31.5 on ProgramBench and 51.7 on Humanity's Last Exam, both behind the Western flagships. And 智东西, which went through the full comparison table, notes that on the longest-horizon agent evaluations — Agents' Last Exam and Agent Arena — Qwen4-Max is strong but still behind Opus 5.5 and GPT-6 Astra, per 智东西's read of the sheet. The pattern is consistent: Qwen4-Max is excellent at bounded professional tasks with a clear deliverable, and merely competitive at open-ended work that runs for hours.

That distinction matters more than the average, because the whole pitch is long-horizon agent work. A model that wins the sub-task and loses the project is the exact failure mode enterprise buyers have spent 2026 learning to test for.

The independent verdict: cheaper than itself, not cheaper than the field

Here is where the story turns, and it turns on cost per task rather than cost per token. Artificial Analysis runs every model through the same evaluation suite and bills it at list price. Here is how the field looked at Qwen4-Max's launch:

  • Gemini 4 Carbon — index 60, $2.00 in / $10.00 out, $5.48 per task
  • Claude Opus 5.5 — index 59, $4.00 in / $20.00 out, $10.04 per task
  • Gemini 4 Hydrogen (xhigh) — index 54, $1.50 in / $7.50 out, $3.16 per task
  • Kimi K3.1 Pro (open weights) — index 46, $2.00 in / $10.00 out, $3.68 per task
  • Qwen4-Max — index 43, $1.50 in / $4.50 out, $3.27 per task
  • Qwen3.8-Max — index 34, $2.00 in / $6.00 out, $10.02 per task

Qwen4-Max's cost per task really did collapse — from $10.02 for Qwen3.8-Max to $3.27, a two-thirds cut that is even steeper than Alibaba's own 62% claim on its Work360 suite. That part of the announcement checks out.

What does not check out is the implication that this makes Qwen4-Max a cheap model. At $3.27 per task it costs eleven cents less than Kimi K3.1 Pro, which scores three points higher, and eleven cents more than Gemini 4 Hydrogen, which scores eleven points higher. Google is selling a better model for less money per task at exactly the tier Qwen is trying to own. That is not a pricing triumph; it is a pricing trap, and it is the same trap we described when Gemini 4 Carbon took the top of the index at 59: the frontier labs now set the price floor, not the challengers.

The per-token framing hides this because output is where the money is spent, and $4.50 per million output tokens looks like a 40% discount to Hydrogen's $7.50. Per task, the discount disappears. Qwen4-Max generated 127 million output tokens across the index — fewer than Hydrogen's 304 million or Opus 5.5's 218 million, but still well above the 78.6 million median for its class. Alibaba made its own model 48% less verbose than its predecessor and it still isn't lean enough to win the bill.

What Alibaba is really selling

None of that makes the launch a failure, because the benchmark scores are not aimed at developers choosing a chatbot. They are aimed at Chinese enterprises buying delivered work.

The Work360 framework is the tell. Alibaba classified professional work into 26 top-level domains, 182 sub-domains and 721 sub-sub-domains, then trained and evaluated against that taxonomy. The demos that follow are all delivery, not conversation: a machining program for a single part that used to take 475 minutes of human programming now takes 24, built with two manufacturing partners; a financial research workflow that calls 90 MCP tools and produces an Excel model plus a slide deck with double sourcing; a physics lesson website that Qwen4-Max shipped in 14 usable versions, each one an expansion of the last.

The most striking claim is the self-referential one. Alibaba says the model autonomously optimised the cloud scheduling systems that serve its own training and inference — cutting data-labelling call costs 10.42%, raising GPU idle capacity utilisation 4.62%, and improving the solve speed on offline CPU rescheduling by more than 14 times against the baseline. Models optimising their own serving stack is either the most credible efficiency story in the announcement or the least verifiable; there is no outside audit of any of it.

Then the science. A 48-hour autonomous run screened 2.7 million candidate molecules and, through four design-synthesise-test cycles, produced a coolant molecule that was actually synthesised, aimed at the data-centre cooling problem. A second run against the NLRP3 target in Alzheimer's research made 454 tool calls, ran 3,239 computational evaluations and produced 24 candidate molecules, the best binding 3.6 times more tightly than the starting point. These are dry-lab results with wet-lab validation claimed for the coolant only, and no independent replication. Treat them as a demonstration of workflow, not a discovery.

The weight question

The most consequential sentence in the launch is the one about what did not ship: the weights.

Qwen4-Max is API-only. Alibaba says smaller Qwen4 models will be open-sourced later, and Bloomberg reported that the Max tier stays closed, per Bloomberg's reporting. For a lab whose reputation was built on Qwen3.8-27B running on consumer GPUs and a 2.4T base model published in the open, that is a reversal at the top of the range, and it is the same pattern as Meituan's LongCat-2.5, which shipped MIT-licensed 1.6T weights and a deployment recipe for 194 Huawei chips. The Chinese open-weight camp is now split between labs that still publish the top of the family and labs that publish everything below it.

That makes the Qwen4-27B release the actual event to watch, not this one. In the meantime, the best open-weight model on the index remains Kimi K3.1 Pro at 46 — twelve points behind Opus 5.5, a gap we argued in the open-weights gap is real, and it is a price list is a structural lag rather than a temporary one. Qwen4-Max does not close it, because you cannot download it.

What skeptics should press on

Three things, in order of how much they would change the conclusion.

First, Work360 and the 62% cost cut are Alibaba's own measurements on Alibaba's own taxonomy and workload mix. The independent index agrees on the direction and disagrees on the magnitude of the advantage once you compare across vendors, which is the only comparison a buyer makes.

Second, the preview build had visible multimodal failures — 智东西's own hands-on with the November 11 preview found inconsistent outputs and physical implausibility on complex tasks, and reports that the shipping version fixed the specific cases. A model that generates a GIF with mismatched lip-sync is not a model whose 48-hour autonomous science run should be taken on faith.

Third, the safety claims are commitments, not measurements. Alibaba says it filtered training data to strip offensive cyber capability and mitigate chemical, biological, radiological and nuclear risk, as South China Morning Post reported. No external red-team results were published alongside the launch, which is now normal for Chinese frontier releases and still a gap that Western enterprise procurement will price in.

What to watch

Watch the timing of the open Qwen4 releases. If a 27B or 125B Qwen4 lands within weeks, Alibaba is repeating the Qwen3.8 playbook — flagship closed, workhorse open — and the open-weight ecosystem gets its next default model. If the small models slip, the Qwen brand's strongest asset starts to erode.

Watch whether the Terminal-Bench and ProgramBench gaps close in a point release. Alibaba's architecture work is aimed exactly at the long agent loops those tests measure, and a 48.5 that becomes a 55 would change the buying conversation more than any price cut.

And watch the price. Qwen4-Max sits at $3.27 per task against Hydrogen's $3.16, and Google has been cutting its own list prices every few weeks. If Alibaba has to follow Google down to hold the tier, the 62% cut it announced today is a down payment rather than the final number. The model is good enough to sell. The price is not yet good enough to win.

Get SiliconNoon in your inbox

One email a day with the AI stories that matter — what happened, the numbers, and why it matters. No spam, unsubscribe in one click.

Read more

Deep Dive — DeepSeek V4.1-Flash-1017: 77% on DeepSWE for 20 cents

Deep Dive — DeepSeek V4.1-Flash-1017: 77% on DeepSWE for 20 cents

DeepSeek refreshed its cheapest model on Friday and the headline number is the one buyers actually care about: 77.0 percent resolution on DeepSWE v1.1, per the model card, at just under $0.20 of compute per task on the company's cost chart. That chart puts Claude Opus 5, GPT-6 Astra and Gemini 3.8 Flash somewhere between $2 and more than $10 for the same benchmark, and none of them above the small Chinese model's score — a gap The Decoder summarized as beating larger rivals on agentic tasks…

Meituan's LongCat-2.5: 1.6T parameters on 194 Huawei chips

Meituan's LongCat-2.5: 1.6T parameters on 194 Huawei chips

Meituan put the weights for LongCat-2.5 on Hugging Face this week under the MIT licence: 1.6 trillion parameters, about 48 billion of them active per token, one-million-token context. The interesting part is not the benchmark table. It is the deployment recipe attached to it — a published way to serve a model that size on accelerators that are not Nvidia's. What actually shipped The model was announced on November 2 and the weights followed on November 9, in two repositories: a full BF16…

Deep Dive — ERNIE 5.5 Flash: 280B params for ¥2 a million tokens

Deep Dive — ERNIE 5.5 Flash: 280B params for ¥2 a million tokens

Baidu released ERNIE 5.5 Flash Preview this morning: 280 billion total parameters, 14 billion of them active per token, a 1-million-token context, and a Chinese list price of ¥2 per million input tokens at peak hours — roughly $0.28. It is the first post-trained version of the base model Baidu published on Hugging Face six days ago, and it is live today through Qianfan, Baidu's cloud, while the weights stay locked until November 25. So the honest answer to the question you are probably typing…

Mistral Large 4 goes open weight — DeepSeek still beats it on cyber

Mistral Large 4 goes open weight — DeepSeek still beats it on cyber

Mistral put the weights of Mistral Large 4 up for download today, three weeks after previewing the trillion-parameter model it nicknamed Le Chonk. Wired's headline calls it a leapfrog over China in the open-weight race, and Mistral's own research post calls it the best open-weight model for coding. The chart inside that same post tells a narrower story: on the cybersecurity index Mistral is showcasing, two Chinese open models score higher. What actually shipped The download is real this…

Deep Dive — Clover 35B: a diffusion LLM at Qwen parity, 680 tok/s

Deep Dive — Clover 35B: a diffusion LLM at Qwen parity, 680 tok/s

DiffuSpace released Clover-35B-A3B-Preview on Tuesday: a 35-billion-parameter mixture-of-experts diffusion language model under Apache 2.0, with weights on Hugging Face and ModelScope, a 128K context window, a thinking mode, and serving paths in SGLang and llama.cpp. It reports 73.6 on SWE-bench Verified, 51.7 on Terminal-Bench 2.1 and 65.3 on τ³ Overall, plus 680 tokens per second on a single GB300 at BF16, batch size one, according to the launch blog. That is the first open-weight diffusion…

Kimi K3.1 Pro: 55 on the AI index, 3 points behind Opus 5.5

Kimi K3.1 Pro: 55 on the AI index, 3 points behind Opus 5.5

Moonshot AI says its new Kimi K3.1 Pro is the best open-weights model anyone has built. The weights are not out yet — the company says they land by November 4, and if that date holds, the gap between the best open model and the best proprietary one shrinks from twelve index points to three. Kimi K3.1 Pro scores 55 on the Artificial Analysis Intelligence Index, fourth overall and three points behind Claude Opus 5.5's 58, according to the evaluator's launch analysis.