Qwen4-Max at $1.50 / $4.50: 2.4 trillion parameters, 71 billion active, a 1M window, and a benchmark story that depends on whose table you read
Alibaba's new flagship is 25% cheaper than its predecessor and competitive on agentic coding, but independent indices put it mid-pack; the model to watch is the open-weight sibling that shares its architecture.

Alibaba moved Qwen4-Max from a two-week preview to general availability on Monday, Beijing time, with international API pricing of $1.50 per million input tokens and $4.50 per million output tokens. That is a quarter below Qwen3.8-Max's $2 and $6, with cached input reads at $0.075 and cache writes at $1.875 for five minutes or $3 for an hour. The preview model retires on November 25.
What is inside
The backbone is the same 2.4 trillion parameters as Qwen3.8-Max, but the active count per token has fallen from 95 billion to 71 billion, and a separate 328 billion-parameter n-gram embedding store has been added. The effect is a model that leans on lookup rather than compute for part of its work, which is the architectural choice behind the lower price. Inputs are text, image and video; the context window is 1 million tokens, with up to 262,144 output tokens in thinking mode and 131,072 without. It serves the OpenAI Chat Completions and Responses formats, the Anthropic Messages format and Alibaba's own DashScope, with function calling, web search and JSON-schema output, and the entry tier allows 5 million tokens a minute.
There are no public weights for the Max model. Alibaba says smaller open-weight Qwen4 models "will soon be released," and has already put out Qwen3.8-Flash-Next, a 125 billion-parameter, 6 billion-active preview of the same architecture under the Qwen Community License.
The two benchmark stories
Alibaba's own table compares Qwen4-Max with Qwen3.8-Max, Claude Opus 5.5, GPT-6.1 Sol, GPT-6 Astra and Kimi K3.1 across 16 tests and shows it leading on four: DeepSWE at 75.7 against Kimi's 75.5, Agents' Last Exam at 46.7 against 43.4, HLAB-Vals at 21.3 and BioMysteryBench at 85.0. The gaps run the other way elsewhere: 48.5 on Terminal-Bench 4.0 against Kimi's 67.2, 31.5 on ProgramBench against Opus 5.5's 65.0, and 54.2 on the Chartography test against GPT-6 Astra's 71.1.
The independent picture is steadier and less flattering. Vals AI ranks the model 11th of 38 at 63.2%, just above Kimi K3.1 Pro and four points above its predecessor, and its Terminal-Bench run matches Alibaba's figure exactly. Artificial Analysis gives it 43 on its intelligence index, 38th of 81, up from 34 for Qwen3.8-Max but well below Claude Fable 5.2 at 65, Gemini 4 Carbon at 60, Opus 5.5 at 59 and GPT-6.1 Sol at 55. On cost per index task the order is GPT-6.1 Sol at $1.27, Qwen4-Max at $3.27, Kimi K3.1 Pro at $3.68 and Opus 5.5 at $10.04.
The honest summary is that Qwen4-Max is a strong agentic-coding model priced aggressively, not a frontier leader. Where it wins, it wins narrowly on tasks that reward tool use; where it loses, it loses by twenty points.
The context that matters more than the scores
Capgemini's survey of 500 executives at companies with more than $1 billion in revenue, published last week, found 59% planning to shift more work to cheaper open-weight models and 53% worried about AI costs, up from 45% a year ago. OpenRouter data in the same report showed open-weight models carrying 72% of traffic but 27% of spend. Anthropic's prospectus names exactly that migration as a risk. A 2.4 trillion-parameter model at $1.50 input is how the migration is marketed, and the release of an open sibling is how it is made safe for buyers who cannot send data to Alibaba Cloud.
That last point is where today's other news intrudes. The BIS rule released on Monday reaches customers who train or supply models for Chinese companies, and Chinese-hosted inference is a data-residency question for any regulated buyer. Qwen4-Max is cheap; using it from a US enterprise is not free of friction.
What to watch
The open-weight Qwen4 releases and their licence terms, which decide whether the architecture can be run on-premises. Third-party agent evaluations on your own tasks rather than the vendor's 16. Whether Alibaba holds $1.50 after the preview promotion period, as its pricing history suggests it will. And how DeepSeek, Moonshot and MiniMax respond: the last round of Chinese price cuts took weeks, not months.
Sources
GeoToolbox launch analysis (pricing, vendor and independent benchmarks) ↗AlphaSignal ↗Alibaba Cloud press release ↗The story so far
- Nov 17, 2026 Google's GTIG finds 'exploit_agent', the first framework that turns a vendor patch into a live exploit with no human in …
- Nov 15, 2026 Microsoft's Lazarus: seven edited neurons out of 12,000 backdoor a coding agent 98.8% of the time and pass every standar…
- Nov 15, 2026 Mirendil's $1 billion Series A at a reported $5 billion, eleven months after founding: what investors are buying when th…
- Nov 15, 2026 AlphaGenome 2: a conditional DNA model trained on 107 million cell profiles finds nearly twice the expression-changing v…
Key terms in this story
- Open-weight model A model whose trained parameters are published for anyone to download and run.
- Tool use (function calling) A model's ability to request that the surrounding software run a function, such as a search or a database query, and return the result.
- Context window The maximum text a model can consider in one request: prompt, documents, conversation and answer combined.
- Evaluation (eval) A test set with known-good answers used to measure a model or system on your own tasks.
- Embedding A list of numbers that represents the meaning of a piece of text, so similar texts have similar numbers.
More in Models & research
All →
Gemini 4 Neon: $3 / $15 per million tokens, 1M context, four thinking levels and a 32% cost-per-task claim against Carbon. Where it sits in Google's four-model lineup and what to test
Google's fourth Gemini 4 release in a month is priced to become the default for paying users and the orchestrator for Hydrogen sub-agents; the benchma…

GPT-6.2 Astra at $8 / $40: OpenAI cuts its flagship price by a fifth, matches Claude Fable 5.2 and ships the first model it rates 'High' for AI self-improvement. What changed and what to test
The headline is a price cut; the system card is the story. OpenAI has released a model it says can materially help build the next one, under added saf…

Microsoft's Lazarus: seven edited neurons out of 12,000 backdoor a coding agent 98.8% of the time and pass every standard eval. What ML teams should change about model provenance
A model file can now carry a trigger-activated payload the way a binary can, and the only reliable detector is a diff against the original weights; th…
Get the brief in your inbox
One email every morning: the stories that mattered in AI, each with a one-line "why it matters" for your role. Free, no ads, unsubscribe any time.
Published daily by AI Briefing. We never share your address.