Alibaba's Qwen4-Max Cuts Reported Task Costs but Trails Rivals on Independent and Agentic Benchmarks
Alibaba says Qwen4-Max cuts costs by around 60% over its predecessor. Artificial Analysis places it 43 on its Intelligence Index, below Opus 5.5, GPT-6.1 Sol and Kimi K3.1 Pro.
Editor's Note ·
- Clarification:
- The article quotes VentureBeat as saying Alibaba reports cutting Work360 task costs "from $4.43 to $1.67, representing approximately 62% savings." VentureBeat's actual wording is "reducing the average cost of a task in its internal Work360 evaluation from $4.43 to $1.67, approximately 62%" — it does not contain the words "representing" or "savings." The figures themselves ($4.43, $1.67, 62%) are accurate.
- Clarification:
- The article quotes VentureBeat as describing token consumption dropping "roughly 24% at the Xhigh reasoning setting and 35% at High." VentureBeat's actual wording uses "approximately," not "roughly": "token consumption falls approximately 24% at the Xhigh reasoning setting and 35% at High on Work360." The figures themselves (24%, 35%) are accurate.
Background
Alibaba introduced Qwen4-Max on November 16, 2026, describing it in a company blog post as a new flagship with “2.4 trillion total parameters, native vision capabilities, and a one-million-token context window.” The post says Qwen4-Max “activates only 71 billion per token—25% fewer activated parameters compared to Qwen3.8-Max at the same 2.4T scale.” A follow-up enterprise-focused post published the next day, November 17, confirms the same parameter counts and context window, and states that the benchmark and cost figures in that particular post “compare Qwen4-Max with Qwen3.8-Max. They are not comparisons with any other provider’s models.” Qwen4-Max launches as a closed, proprietary model; according to VentureBeat and The Decoder, smaller Qwen4-series models are planned for release as open weights, though the flagship itself is not.
The release is the latest in a string of frontier-model announcements this month that pair cost or efficiency claims with benchmark tables, including Google’s Gemini 4 Hydrogen, where Artificial Analysis — a third-party benchmarking site that scores models on its own Intelligence Index — scored Hydrogen 51 against 50 for Gemini 4 Argon, a result the Herald reported alongside Google’s own, larger efficiency claims. The Qwen4-Max release follows the same pattern: a vendor’s own efficiency and task-performance claims, checked against an independent benchmarking site’s aggregate score, with the two telling somewhat different stories about how the model actually ranks against competitors.
Current situation
Alibaba’s own framing centers on cost. The November 16 post says Qwen4-Max “outperforms its predecessor, Qwen3.8-Max, overall while using 35% fewer tokens at High effort reasoning mode and reducing the average cost per task by around 60%,” measured on Work360, which the post describes as “a proprietary benchmark evaluating performance on challenging industry tasks across 721 specialized fields.” According to VentureBeat, Alibaba also reports cutting average task costs on Work360 “from $4.43 to $1.67, representing approximately 62% savings,” with token consumption dropping “roughly 24% at the Xhigh reasoning setting and 35% at High.” Standard API pricing, confirmed by both VentureBeat and The Decoder, is $1.50 per million input tokens and $4.50 per million output tokens, with Alibaba’s own post listing a cached-input rate of $0.075 per million tokens. Alibaba’s post also says the model “ranked 6th on Code Arena, a real-world coding leaderboard,” though that figure, like the Work360 numbers, comes from Alibaba’s own announcement.
VentureBeat separately reproduced a wider benchmark chart that Alibaba published alongside the launch, covering six evaluations — DeepSWE v1.1, Terminal-Bench 4.0, ProgramBench, APEX-Agents v1.1, AutomationBench-AA and ChartoGraphy — against Qwen3.8-Max, Claude Opus 5.5, GPT-6.1 Sol and Kimi K3.1. On that chart, Qwen4-Max scores higher than Qwen3.8-Max on every listed benchmark. Against the other three models, its standing varies by benchmark: on APEX-Agents v1.1 it beats GPT-6.1 Sol (72.1 to 60.0) but trails Opus 5.5 (73.5); on AutomationBench-AA it beats GPT-6.1 Sol (72.5 to 64.9) but trails Kimi K3.1 (73.3). It trails more sharply on Terminal-Bench 4.0, where it scores 48.5 against Kimi K3.1’s 67.2, and on ProgramBench, where it scores 31.5 against Opus 5.5’s 65.0. Critically, VentureBeat is explicit about where these numbers come from: “These figures reproduce Qwen’s chart, including its Kimi label. They are not independently rerun results. Alibaba’s footnotes attribute some scores to outside evaluators, including Mercor and Artificial Analysis, and describe specific settings or dataset versions for others.” In other words, the comparison table that frames Qwen4-Max against its rivals is Alibaba’s own selection and presentation, not a benchmark run end-to-end by an outside party.
Artificial Analysis’s own Intelligence Index, which the site runs independently of the vendors it scores, places Qwen4-Max well behind several of the same rivals named in Alibaba’s chart. VentureBeat reports the index chart at launch: “Qwen4-Max at 43 on its Intelligence Index, up from 34 for Qwen3.8-Max. Kimi K3.1 Pro scores 46; GPT-6.1 Sol and Claude Sonnet 5.5 score 55; Opus 5.5 reaches 59; Gemini 4 Carbon scores 60; and Claude Fable 5.2 leads this displayed group at 65.” Artificial Analysis’s own model page for Qwen4-Max describes the 43 score as “below average among other models in a similar price tier,” and separately confirms the $1.50/$4.50 per-million-token pricing along with a blended rate of $0.80 per million tokens using the site’s standard 7:2:1 cache-hit/input/output ratio.
The Decoder’s reporting adds a cost-per-task framing specific to the Intelligence Index itself: “Artificial Analysis puts the cost of Qwen4-Max at $3.27 per Intelligence Index task. That’s about a third of Qwen3.8-Max’s $10.02 and roughly 11 percent less than Kimi K3.1 Pro’s $3.68.” The Decoder also describes Qwen4-Max as ranking “second behind Kimi K3.1 Pro” on the Intelligence Index, but its own numbers make clear this is a ranking among Chinese models specifically — the outlet calls Kimi K3.1 Pro “the top Chinese model at 46 points,” three ahead of Qwen4-Max’s 43 — not a claim that Qwen4-Max is the second-highest-scoring model overall; VentureBeat’s fuller chart shows GPT-6.1 Sol, Claude Sonnet 5.5, Opus 5.5, Gemini 4 Carbon and Claude Fable 5.2 all scoring higher.
On benchmarks outside Alibaba’s own chart, The Decoder reports a mixed picture that does not uniformly favor either side: “On long-horizon agent tasks like Agents’ Last Exam v1.1, Qwen4-Max scores 46.7 points, the best result in the comparison group. It trails Opus 5.5 and the GPT models on Humanity’s Last Exam. On HealthBench Professional, it falls behind the GPT models but edges past Opus 5.5.” The outlet summarizes Alibaba’s own position as: “Alibaba says Qwen4-Max leads on agent and work tasks, while Opus 5.5 and the GPT models come out ahead on Humanity’s Last Exam.”
Implications
The following points restate what the sources say; they are not predictions.
The gap between Alibaba’s chart and Artificial Analysis’s index illustrates a split in how “improvement” and “competitiveness” get presented in frontier-model launches this month. Alibaba’s own numbers — the ~60% cost reduction on Work360, the Code Arena placement, and the six-benchmark comparison chart VentureBeat reproduced — are drawn from a benchmark Alibaba built itself (Work360) and a chart VentureBeat says was not independently rerun. Artificial Analysis’s Intelligence Index, run by a third party on all the models it lists, places Qwen4-Max well below Opus 5.5, Gemini 4 Carbon, Claude Fable 5.2, GPT-6.1 Sol and Claude Sonnet 5.5, and only modestly below Kimi K3.1 Pro. Both sets of figures are accurate to what they measure; they simply measure different things, one of them Alibaba’s own task suite, the other a standardized index applied the same way across vendors.
Cost and capability are reported separately, and the two don’t move together in the same direction across the sources. Qwen4-Max is cheaper than Qwen3.8-Max by Alibaba’s own account, and Artificial Analysis’s cost-per-Intelligence-Index-task figure ($3.27) is lower than both Qwen3.8-Max’s ($10.02) and Kimi K3.1 Pro’s ($3.68). But a lower cost per task on an index where the model itself scores below several of its peers does not establish that it is a better value at matched capability — none of the cited sources make that comparison directly, and the Herald has not computed one.
Where Qwen4-Max leads and where it trails depends heavily on which benchmark is asked. It leads its own comparison group on Agents’ Last Exam v1.1 per The Decoder, edges past Opus 5.5 on HealthBench Professional, and beats Qwen3.8-Max across the board on Alibaba’s six-benchmark chart. It trails substantially on Terminal-Bench 4.0 and ProgramBench in that same chart, and trails Opus 5.5 and GPT-6.1 Sol-class models on Humanity’s Last Exam and the aggregate Intelligence Index. No single cited figure settles whether Qwen4-Max is a stronger or weaker model than its named rivals; the sources instead describe a model that is cheaper and competitive on some tasks, and clearly behind on others.