Qwen4-Max Is Alibaba's Best Model and Its Weights Stay Closed. Artificial Analysis Scores It 43, and GPT-6.1 Sol Scores 12 Points Higher for Less Than Half the Cost per Task.
Alibaba launched Qwen4-Max on November 16 as qwen4-max on QwenCloud's API and Token Plan, QwenWork, and QwenStudio: a 2.4 trillion parameter mixture of experts with 71 billion active by Alibaba's count, $1.50 per million input tokens and $4.50 output in Singapore and the international region, a 1 million token context, and thinking plus answer capped at 262,144 tokens. The weights are not released; smaller open Qwen4 models are promised with no date. Artificial Analysis scores it 43 at xhigh on Intelligence Index v5.0.1, 38th of 81, nine points above Qwen3.8 Max, at $3.27 per Index task on 127 million output tokens. GPT-6.1 Sol scores 55 at $1.27, and Kimi K3.1 Pro 46 at $3.68. Alibaba's coding claims are vendor numbers, and its own chart shows Terminal-Bench 4.0 losses. The two-week preview ID still retires November 25. We read the launch post, the Model Studio doc, and AA.
By T.W. Ghost
Alibaba launched Qwen4-Max today, November 16, 2026, five days after a two-week preview that had no price, no benchmark, and no way in outside a subscription. The GA model is qwen4-max, on QwenCloud's API and Token Plan and in QwenWork and QwenStudio, with pay-as-you-go now supported. Alibaba calls it a 2.4 trillion parameter mixture of experts with 71 billion active, 25% fewer active parameters than Qwen3.8-Max. Price is $1.50 per million input tokens and $4.50 output in Singapore and the international region, $1.413 and $4.24 in Beijing, Virginia, Frankfurt, Hong Kong, and Tokyo, with implicit cache hits at $0.075 and 1 million free tokens for 90 days in Singapore. Context is 1 million tokens; thinking plus answer is capped at 262,144, down from the preview's 393,216; the default effort is high, down from the preview's xhigh, and medium maps to high. No Batch, no fine-tuning, no model deployment.
Two facts frame the rest. The weights are closed, a break from the Qwen line's habit, with smaller open Qwen4 models promised "soon" and no date. And Artificial Analysis scored it the same day: 43 at xhigh on Intelligence Index v5.0.1, 38th of 81, nine points above Qwen3.8 Max at 34, at $3.27 per Index task on 127 million output tokens, which AA calls "very verbose" against a 78.6 million median. On the same board GPT-6.1 Sol scores 55 at $1.27 per task and Kimi K3.1 Pro 46 at $3.68; AA puts Qwen4-Max's cost per task about 11% below Kimi's. Alibaba says Qwen4-Max beats Fable 5.1 and Kimi K3.1 Pro at coding. Its own chart, relayed by The Next Web, puts it at 48.5 on Terminal-Bench 4.0 against Opus 5.5's 59.6 and Kimi K3.1's 67.2. We read Alibaba's launch post, the Model Studio doc, AA's model page, and AA's post. Here is what held up.
What You Are Calling
| qwen4-max (November 16, 2026) | |
|---|---|
| Where | QwenCloud API and Token Plan; QwenWork; QwenStudio; pay-as-you-go supported |
| Size | 2.4T total MoE; 71B active; 25% fewer active than Qwen3.8-Max (vendor) |
| Weights | Closed; smaller Qwen4 open models "soon," no date |
| Price | $1.50 / $4.50 per 1M (Singapore, International); $1.413 / $4.24 (Beijing, Virginia, Frankfurt, Hong Kong, Tokyo); implicit cache hit $0.075; 1M free tokens for 90 days (Singapore) |
| Context and output | 1M context; thinking plus answer capped at 262,144 tokens (preview: 393,216) |
| Effort | low, high, xhigh; default high (preview default xhigh); medium maps to high |
| Not supported | Batch; fine-tuning; model deployment |
| AA Intelligence Index v5.0.1 | 43 at xhigh; 38th of 81; Qwen3.8 Max 34 |
| AA cost and tokens | $3.27 per Index task; 127M output tokens (median 78.6M, "very verbose"); time per task not available |
| Preview ID | qwen4-max-preview still retires November 25, 10 AM Central |
Closed weights are the news. Every Qwen flagship we have covered shipped or promised weights. This one does not, and the promise attached is for smaller models at an unstated date.
Twelve points below Sol at more than twice the cost per task. Forty-three against fifty-five, and $3.27 against $1.27. The per-token price is not high; the token count is, and the Index bill follows the tokens.
Very verbose, by AA's label. One hundred twenty-seven million output tokens against a median under eighty million, with the time column empty on AA's page.
The coding win is selective. DeepSWE v1.1 at 75.7 against Opus 5.5's 74.2 is Alibaba's chart; so is Terminal-Bench 4.0 at 48.5 against Opus 5.5's 59.6 and Kimi K3.1's 67.2, and the same chart cannot support "beats Kimi K3.1 Pro at coding" and show that row.
The Benchmarks, Sorted by Who Reported Them
| Claim | Number | Who |
|---|---|---|
| Intelligence Index v5.0.1 | 43 at xhigh; 38th of 81; +9 over Qwen3.8 Max | Independent (Artificial Analysis) |
| Cost per Index task | Qwen4-Max $3.27; Kimi K3.1 Pro $3.68; GPT-6.1 Sol $1.27; Qwen3.8 Max $10.02 | Independent (AA) |
| Output tokens | 127M vs 78.6M median | Independent (AA) |
| Code Arena sixth; Work360 35% fewer tokens, about 60% lower cost per task ($4.43 to $1.67) | Vendor (Alibaba; via TNW) | |
| DeepSWE v1.1 75.7 vs Opus 5.5 74.2; Terminal-Bench 4.0 48.5 vs Opus 5.5 59.6 and Kimi K3.1 67.2 | Vendor chart (via FourWeekMBA and TNW; not read in the original) | |
| Beats Fable 5.1 and Kimi K3.1 Pro on coding | Vendor claim (via Bloomberg and TNW) |
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| Launched Nov 16, 2026; model id; surfaces; pay-as-you-go | Verified (Alibaba blog; Model Studio doc) |
| 2.4T total, 71B active | Vendor number |
| Closed weights; smaller open models "soon" | Verified (vendor) |
| Prices by region; cache; free tokens | Verified (Model Studio doc) |
| Context, output cap, effort levels, unsupported features | Verified (Model Studio doc) |
| AA score, rank, cost, tokens | Verified (AA model page; AA post) |
| Coding claims and chart | Vendor |
| Preview retires Nov 25 | Verified (preview doc) |
Who Should Use It
Use Qwen4-Max if you are already on QwenCloud and want Alibaba's strongest model with a per-token price under Sol's, and you have measured your own token counts, because the Index says it talks a lot.
Skip it for open weights, which it does not have, and skip the coding headline until a board other than Alibaba's chart carries it; migrate anything on the preview ID before November 25.
The frontier is elsewhere: the open-weights leader at 46 is Kimi K3.1 Pro, the price-performance row is GPT-6.1 Sol, and the top of the board is Claude Fable 5.2 at 65. Our model comparison has the current standings side by side.
Sources
- Alibaba Cloud: Alibaba unveils Qwen4-Max
- Alibaba Cloud Model Studio: qwen4-max
- Artificial Analysis: Qwen4-Max model page
- Artificial Analysis on X: Qwen4-Max
- The Next Web: Alibaba launches Qwen4-Max flagship AI model without open weights
*Not sure whether a closed flagship at an open-model price fits your stack? Take the free 2-minute quiz and get matched. Then read about the two-week preview this replaces: Qwen4-Max Preview.*