Alibaba's Qwen 3.7 Max Sets New Benchmarks for AI Agent Models With 1M-Token Context
Alibaba launched Qwen 3.7 Max, a closed-weight frontier model scoring 60.6 on SWE-Pro and 92.4 on GPQA Diamond, priced at $2.50/$7.50 per million tokens. It executes 1,000+ autonomous tool calls per task and introduces native extended-thinking mode designed for long-horizon agentic workflows.
Alibaba’s Qwen team shipped Qwen 3.7 Max on May 19 — a closed-weight frontier model that benchmarks above every competing system on agentic coding and long-horizon reasoning tasks.
The numbers are hard to dismiss: 60.6 on SWE-Pro (the software engineering benchmark), 69.7 on Terminal-Bench 2.0, and 92.4 on GPQA Diamond. Those scores place it above the current frontier tier on tasks that specifically test whether a model can hold a long plan together without collapsing into hallucination.
What makes this one different
Qwen 3.7 Max is built around a 1-million-token context window and a native extended-thinking mode that activates when the model determines a problem requires multi-step planning. It’s not bolted on — the extended-thinking architecture is the default inference path for complex tasks, not a switchable flag.
The team demonstrated a single task run where the model executed over 1,000 autonomous tool calls sequentially: reading files, running tests, checking logs, and writing patches without human checkpoints. That’s a qualitative jump from models that handle 50-100 tool calls before coherence degrades.
Inference speed is described as 10x faster than prior Qwen releases, attributed to a new compiler-level optimization running on Alibaba’s custom chip platform. On third-party providers the speed improvement is smaller, but latency on DashScope (Alibaba’s own API) is noticeably lower than comparable closed models.
Pricing and access
At $2.50 per million input tokens and $7.50 per million output tokens, Qwen 3.7 Max is API-only — no open weights. It’s available on DashScope, OpenRouter, and Together AI. That pricing puts it below GPT-5.5 Instant for input cost while offering a longer native context window.
Alibaba has not announced when or whether open weights will follow. The previous Qwen releases followed a pattern of closed API first, then a community-released version. There’s no timeline for that here.
Implications for the agent market
The benchmark trifecta — coding (SWE-Pro), terminal ops (Terminal-Bench), and graduate-level reasoning (GPQA Diamond) — covers the three core capabilities that current enterprise agent frameworks depend on. A model that scores well across all three at this pricing tier reshapes the cost structure for agentic SaaS products being built right now.
For developers building agent pipelines, the 1M context window and extended thinking mode are the practical differentiators. Holding full codebase context without chunking, then maintaining coherent reasoning through 1,000+ tool steps, removes a class of engineering hacks that most current pipelines depend on.
The closed-weights decision remains a friction point for teams that need to self-host for compliance reasons. But for API-first development the benchmark profile and pricing make Qwen 3.7 Max worth serious evaluation against the current defaults.