Qwen3.7 Max is Alibaba's flagship, and as of July 2026 it was the highest-ranked Chinese model on the Artificial Analysis Intelligence Index at 56.6 (global #5, domestic #1): a 1M token context, 35 hours of unattended autonomous execution, and an API at $2.50 in / $7.50 out per million tokens, roughly a third of what Claude Opus 4.7 charges on output. If you've been writing off Chinese labs as "good enough for the domestic market," this is the model that should make you stop. Date stamp: the numbers are the July 2026 snapshot, taken before the fall index reshuffle; treat them as history, not live standings.
The Chinese workhorse, versioned
Vendor: Alibaba Cloud, built by the Qwen team at the Tongyi Lab. Released May 20, 2026 at the Alibaba Cloud Summit, as the top of the Qwen3.7 family (which also ships a cheaper, vision-capable Qwen3.7-Plus). The official positioning, translated: "a new-generation flagship for the agent era." It runs on a trillion-parameter MoE architecture, carries a 1M token context window, and supports dual thinking/non-thinking modes so you can trade depth for latency.
Two details worth flagging. First, it went from public unveiling to live API on Bailian (Alibaba Cloud's model platform) in two days (May 22). Second, it's API-only. Previous Qwen models shipped open weights; this one doesn't. The open-source community grumbled, and fairly so. But the capability leap is the reason they closed it.
What 35-hour autonomy buys
Numbers first: Artificial Analysis, Arena, and Alibaba's own release notes, July 2026 — not the launch keynote.
- Artificial Analysis Intelligence Index: 56.6, a 4.8-point jump over Qwen3.6 Max Preview (51.8), and ahead of Gemini 3.5 Flash (55.3).
- MMLU-ProX: 87%, #1 globally, ahead of Claude Opus 4.5 (85.7%). Multilingual professional knowledge, and Qwen3.7 Max tops the whole board.
- GPQA Diamond: 92.2. Frontier-grade reasoning.
- Terminal-Bench 2.0 Terminus: 72.3. Domestic SOTA, multiple sub-scores (SWE-Pro 60.4, SWE-Multilingual 78.4, SciCode 52.7) that set global bests.
- MCP-Mark: 64.6. Deep-Planning: 63.1. SpreadSheetBench: 84.5. General-agent and office-automation benchmarks, all top-tier.
- Kernel Bench L3: 2.37x median speedup, 100% pass rate. It optimizes GPU kernels on its own.
- WMT24++ 85.9, MAXIFE 89.3, IFBench 81.2. Global lead in multilingual translation and instruction-following.
The headline demo, though, isn't a benchmark. On a brand-new chip platform, Qwen3.7 Max ran for 35 hours straight, made 1,158 tool calls, and self-evolved a GPU kernel to 10x the original inference speed, with no human in the loop. That's not a score; that's a proof of existence for long-horizon agents. Most frontier models tap out well before that.
Caveats: it's text-only. No vision. (Vision goes to Qwen3.7-Plus.) Knowledge cutoff is mid-2025, no real-time web access. And past 500K tokens the middle of the context starts to soften. The million-token window is real, but it isn't magic.
The Bill, and the Geography Catch
API on Alibaba Cloud Bailian, list price: 12 yuan per million input tokens, 36 yuan per million output tokens. The 2026 promo knocks that to 50% off: 6 yuan in, 18 yuan out. Cache hits drop to 0.6 yuan/MTok; batch inference stacks another 50% off. One million tokens free for new accounts. There's also a Token Plan monthly subscription for teams that want flat-rate billing with priority scheduling.
In USD terms (international deployment): $2.50 in / $7.50 out per million tokens. Compare GPT-5.4 (list price around $17.50/MTok as of June 2026) with Claude Opus 4.7 at ~$15 in / $25 out. On output, Qwen3.7 Max is roughly a third of Opus 4.7's rate, before the promo and cache discounts even kick in. For high-volume Chinese workloads, the gap gets absurd fast.
The catch for US/EU users: it runs through Alibaba Cloud's infrastructure. If data residency or vendor geography matters to your compliance team, that's a real consideration, not a footnote.
Where the Ranks Land
As of the July 2026 snapshots:
- Artificial Analysis Intelligence Index: #5 globally, #1 among Chinese models (56.6).
- MMLU-ProX: #1 globally (87%).
- Code Arena: #4 at 1541 Elo, the first Chinese model ever to crack that table, slotting between Claude Opus 4.6 Thinking and Opus 4.6.
- Arena math: #7. Software & IT: #9. Expert arena: #9.
- Arena overall (text): #13. Solid but not top-tier.
- Vision Arena: #16. Highest-ranked Chinese lab, though remember Max itself is text-only; this reflects the broader Qwen3.7 family.
The pattern is telling. On pure knowledge and reasoning (MMLU-ProX, GPQA), it's at or near the global top. On the Arena chat vote, which rewards vibe and conversational polish, it's mid-pack. On coding and agentic tasks, it's the strongest Chinese model by a clear margin and genuinely competitive globally. Read that as: a serious workhorse, not a charming conversationalist.
Qwen3.7 Max Against the Field
Priced and timed to the same July snapshot:
| Model | AA Index | Pricing (in/out per MTok) | Best at | |-------|----------|---------------------------|---------| | Claude Fable 5 | ~60 (#1) | $10 / $50 | Long reasoning, English prose | | GPT-5.6 Sol | ~58.9 (#2) | $5 / $30 | Math, tools, balanced | | Claude Opus 4.7 | 57.3 | $15 / $25 | Agentic, long docs | | Qwen3.7 Max | 56.6 (#5, #1 CN) | $2.50 / $7.50 | Chinese, cost, autonomy | | Gemini 3.5 Flash | 55.3 | $1.50 / $9 | Speed, multimodal | | DeepSeek V4 | high GPQA | ~$0.14 / $0.28 | Pure volume |
On Artificial Analysis's separate coding and agentic indices (June 2026), the pecking order is GPT-5.5 (Coding 74.9, Agentic 74.1), Claude Opus 4.8 (Coding 56.7, Agentic 77.8), then Qwen3.7 Max (Coding 50.1, Agentic 66.6), the highest Chinese model on both, ahead of Kimi K2.6 and GLM-5.0.
Plainly: Qwen3.7 Max is not beating Fable 5 or GPT-5.6 on raw intelligence. It's roughly 2-4 points behind on the AA Index, and the coding-index gap to GPT-5.5 (50.1 vs 74.9) is real. But it costs a fraction of what they cost, and on agentic persistence (35 hours, 1158 tool calls) it's in a class of basically one. Developer Paul Couvert's verdict after hooking it to Hermes Agent and OpenCode: it can "basically replace GPT-5.5 and Opus 4.7" for a lot of dev work. He's a working dev with receipts, by his own demos. In one Atomic Chat test, building a self-training Tetris AI, Qwen3.7 Max did it for $1.32 in tokens and outperformed both Western flagships.
Assign it like a workhorse
- Chinese-first content: legal contracts in Chinese, financial-report analysis, long-form drafting. More natural than any Western frontier model, at a tenth of the price.
- Long-horizon agents: anything you'd otherwise babysit. The 35-hour autonomous run isn't a parlor trick.
- Cost-bound dev work: high-volume coding tasks where Opus-level prices would sink the budget.
- Enterprise in China: compliance, data residency, native integration with Alibaba Cloud's 400+ ecosystem services.
Don't use it for:
- Multimodal / vision tasks. It's text-only; use Qwen3.7-Plus.
- Real-time web-grounded Q&A. No live search, mid-2025 cutoff.
- The absolute hardest English reasoning. Fable 5 and GPT-5.6 still edge it. If that last 3% matters, pay for it.
The mental model: Qwen3.7 Max is the workhorse you run all day, not the specialist you call for the one impossible problem.
The same job, priced in credits
I cover Chinese labs for x-rush.ai, where none of this naming survives: a text job goes in, routing happens upstream, and the bill stays a flat 0.5 credits a unit. That design is why no routing claim belongs in a model review: whether any given lab is in the rotation is an implementation detail the product deliberately does not promise. Text generation bills the same flat rate whether the engine behind the task was a frontier flagship or something far cheaper, and the assistance pool around chat adds 0.5 credits per ten successful replies.
Verdict: built for the long shift
At the July snapshot, Qwen3.7 Max was the strongest Chinese model shipping: #5 globally and #1 domestically on the AA Intelligence Index, #1 on MMLU-ProX, the first Chinese model in Code Arena's top tier, and the only frontier model publicly demonstrated at 35 hours of unattended autonomous work. It is also text-only, mid-pack on chat-vibe Arena, and a few points behind Fable 5 and GPT-5.6 on raw intelligence. None of that is a contradiction. It is the cheap, agent-persistent one, and the fall index refresh did not change its job.
Run the workhorse all day. Or save it for the one job nobody else finishes.
AI Chat is that trade in practice: the job arrives in plain language, the model choice stays inside.