AI Aulis

AI NEWS IN ENGLISH · 17 November 2026

The most important AI news every morning. AI writes and checks the stories against the original sources – check important facts in the sources.

← Issue 17 November 2026Paper

Artificial Analysis: Qwen4-Max scores 43 points, three fewer than Kimi K3.1 Pro

AI-generated contentWritten by AI

huggingface.co, artificialanalysis.ai, the-decoder.com, heise.de, aitimes.com, vals.ai, qwen.ai · 7 sources · Sources published: 3 August–16 November 2026 · Some source publication dates are unknown

In Artificial Analysis’s evaluation, Alibaba’s Qwen4-Max scored nine points more than its predecessor, and an index task costs about a third of the predecessor’s price. Western frontier models remain well ahead in the index, and the model is slower than models in the same price bracket.

Independent benchmarking service Artificial Analysis published its results for Alibaba’s new frontier model Qwen4-Max on 16 November. The model became generally available through the API the same day, heise online reported.

Artificial Analysis’s Intelligence Index combines nine tests that measure skills such as agentic coding, reasoning and knowledge. At its highest reasoning effort, Qwen4-Max scored 43 points, nine more than its predecessor Qwen3.8-Max. It improved in every test, heise said.

Alibaba’s Qwen team wrote in a blog post the same day that Qwen4-Max is a mixture-of-experts (MoE) model. It has 2.4 trillion parameters but uses only 71 billion of them to process each token, a quarter fewer than its predecessor. For long inputs, three layers compress the earlier text and a fourth layer extracts what matters from it.

Close to the leading Chinese model, behind Western ones

Among Chinese models, Moonshot AI’s Kimi K3.1 Pro leads the index with 46 points. Western frontier models remain well ahead, heise said. The table shows index scores and average cost per index task.

Model Index score Cost per task
Claude Fable 5.2 65 –
Gemini 4 Carbon 60 –
GPT-6.1 Sol (Max) 55 $1.27
Claude Sonnet 5.5 55 –
Kimi K3.1 Pro 46 $3.68
Qwen4-Max 43 $3.27
Qwen3.8-Max 34 $10.02

Source: Artificial Analysis, as reported by heise online and The Decoder; – = figure not given.

Although its tasks cost less than its predecessor’s, Qwen4-Max is not the cheapest option: GPT-6.1 Sol scores 12 points more at less than half the cost.

Cheaper tasks, but slow runs

The Decoder said the savings come from three things. Input and output tokens are a quarter cheaper than for the predecessor, at $1.50 and $4.50 per million tokens. Cached input costs 70 per cent less than before. The model also uses fewer tokens: heise reported that across the whole index it consumes 32 per cent fewer input tokens and produces about 105,000 output tokens per task, against 171,000 for its predecessor. In agentic tasks, 85–90 per cent of the cost comes from input because earlier information is sent to the model again and again, Artificial Analysis said.

On its model page, Artificial Analysis describes Qwen4-Max as notably slow and very verbose. An index task takes 32.1 minutes on average, against a median of 13.3 minutes for the comparison group.

A fifth of legal tasks solved

Vals AI runs a benchmark published by Harvey in which a model does legal work as an agent using file tools. In the table Vals updated on 11 November, Qwen4-Max solved 21.25 per cent of tasks and ranked seventh out of 39 models. Gemini 4 Argon led with 27.08 per cent. Kimi K3.1 Pro solved 11.67 per cent of tasks and Claude Opus 5.5 9.17 per cent.

Two language models grade the answers separately. Each counts a task as solved only if every grading criterion is met, and the result is the average of the two graders’ scores. Qwen4-Max met 95.69 per cent of individual criteria.

Qwen4-Max is available through the QwenCloud API and in Alibaba’s QwenWork, Qoder and QwenStudio services. It can also be used via OpenRouter and Vercel AI Gateway, heise reported. Alibaba is not releasing the model weights. The company plans to release smaller Qwen4 models later as open-weight models but has not given a timeline.

Previous coverage

Report an error in this story