Alibaba's Qwen3.7-Max Ran Autonomously for 35 Hours Beating Claude at One-Sixth the Price

Alibaba's Qwen team quietly did something remarkable: they shipped a proprietary model that ran autonomously for 35 hours, fired 1,158 tool calls without a human in the loop, and delivered a 10× speedup on a GPU kernel the model had never encountered during training. Then they priced it at $2.50 per million input tokens , roughly one-sixth the cost of Claude Opus 4.7. That model is Qwen3.7-Max, and it is the most credible challenge yet to the idea that frontier agentic AI belongs exclusively to OpenAI and Anthropic.

Built for the Agent Era, Not the Chat Era

Announced on May 20, 2026, at the Alibaba Cloud Summit, Qwen3.7-Max is a proprietary, text-focused reasoning model engineered specifically for the agentic era. The Qwen team describes it as a "versatile agent foundation" rather than a general-purpose chat model , a distinction that shows up clearly in the benchmark profile. It wins where it matters for production agents: long-horizon execution, MCP tool orchestration, and real-world software engineering. It does not lead on every raw intelligence metric, and Alibaba is not pretending otherwise.

Three things define the release:

• 1-million-token context window , roughly 2,000 pages of text or a full mid-sized codebase in a single request, up from 262K on the previous generation
• Cross-harness generalization , the model was trained to perform consistently across different agent scaffolds, not just the one it was evaluated on
• Long-horizon autonomy , the 35-hour autonomous session is the longest publicly documented agent run from any major lab

The 35-Hour Kernel Optimization Run

The flagship demonstration is worth understanding in detail, because the details are what make it significant. The Qwen team handed the AI a tough coding problem , optimizing GPU code , giving it just the instructions and a way to test its work, and then stepped back. From there, Qwen3.7-Max worked autonomously, writing code, running tests, finding bottlenecks, and redesigning the code. It looped through this process over a thousand times.

It executed 1,158 tool calls, ran 432 kernel evaluations, diagnosed failures, and achieved a 10.0x geometric mean speedup , all without human intervention. The hardware was Alibaba's own ZW-M890 PPU, a chip the model had never seen during training. No profiling data, no hardware documentation, no example kernels. The model had to navigate an unfamiliar architecture and maintain a coherent optimization strategy across the entire context window for over a day.

For comparison, other top-tier models that were tested on the same task maxed out at 7.3× speedup (GLM-5.1), 5.0× (Kimi K2.6), and 3.3× (DeepSeek V4 Pro) , and all of them tapped out far earlier. Qwen3.7-Max was still finding meaningful optimizations past the 30-hour mark.

One caveat worth stating clearly: this was Alibaba's own internal benchmark, on Alibaba's own hardware. Independent reproduction is just beginning. Treat the 35-hour number as a strong directional signal, not a guarantee your agent will run unattended for a day and a half on day one.

Benchmark Numbers That Actually Matter

The benchmark profile is unusual for a frontier model , it does not lead on every raw intelligence metric, but it wins on the metrics that matter for production agents.

On coding agent benchmarks:

• Terminal-Bench 2.0-Terminus: Qwen3.7-Max scores 69.7, ahead of DeepSeek-V4-Pro Max (67.9), Opus-4.6 Max (65.4), and K2.6 Thinking (66.7). This benchmark tests autonomous terminal-based software engineering with a 5-hour timeout and 12 CPU cores.
• SWE-Pro: It scores 60.6, the highest in the comparison table, ahead of K2.6 Thinking (59.5) and DS-V4-Pro Max (59.0).
• SWE-Verified: It scores 80.4 against Opus-4.6 Max's 80.8 and DS-V4-Pro Max's 80.6. The gap is small , this is the one coding benchmark where it does not lead.
• SWE-Multilingual and SciCode: It leads the field on both.

On general agent benchmarks:

• MCP-Mark: It scores 60.8, ahead of GLM-5.1 Thinking's 57.5 and Opus-4.6 Max's 56.7.
• MCP-Atlas: It scores 76.4, edging out Opus-4.6 Max's 75.8.
• SpreadSheetBench-v1: It scores 87.0, second only to Opus-4.6 Max's 89.3.

On STEM reasoning:

• GPQA Diamond: Qwen3.7-Max achieves leading results on GPQA Diamond (92.4 vs. Opus-4.6's 91.3).
• HMMT 2026 Feb: It scores 97.1, the highest in the table, ahead of Opus-4.6 Max (96.2) and DS-V4-Pro Max (95.2).
• Apex benchmark: It scores 44.5, well ahead of DS-V4-Pro Max's 38.3 and Opus-4.6 Max's 34.5.

On multilingual performance:

• WMT24++: It scores 85.8, confirming top-tier multilingual understanding and translation quality.
• PolyMATH: It scores 86.5, well ahead of Opus-4.6 Max's 80.2 and K2.6 Thinking's 82.7.

Qwen3.7-Max also reports the lowest hallucination rate among frontier models at 22.9%.

The Real Breakthrough: Cross-Harness Generalization

Most agent models are trained and evaluated on a specific scaffold, which means their benchmark numbers can reflect harness-specific shortcuts rather than genuine problem-solving. Qwen3.7-Max is designed to avoid this through a training infrastructure that decouples Task, Harness, and Verifier into three independent components that can be freely recombined.

In practice, this means the model was trained on identical tasks paired with diverse harnesses and verifiers, forcing it to learn generalizable strategies. Notably, these scores are drawn from a wide variety of agent scaffolds. The benchmark results reflect this: Qwen3.7-Max performs consistently whether deployed through Claude Code, OpenClaw, Qwen Code, or custom tool-use frameworks.

For teams building agent systems, this matters because it means Qwen3.7-Max can serve as a drop-in backbone without requiring framework-specific tuning. That is a real operational advantage over models that perform well only in their native scaffold.

The model also supports native integration with Model Context Protocol (MCP) , the standardized way to connect models to external tools and data sources. Its MCP scores are the highest published by any model, which is particularly meaningful because MCP benchmarks are harder to game through training data contamination.

Using Qwen3.7-Max with Claude Code

One of the more practical details in this release: Qwen3.7-Max supports the Anthropic API protocol natively, enabling seamless integration with tools like Claude Code. Three environment variables are all you need:

export ANTHROPIC_BASE_URL=https://dashscope-intl.aliyuncs.com/apps/anthropic
export ANTHROPIC_API_KEY=YOUR_DASHSCOPE_KEY
export ANTHROPIC_MODEL=qwen3.7-max
claude

Tool use, file edits, and subagent features all work because Qwen3.7-Max implements the same protocol Claude Code expects. The deeper Anthropic-specific agent features (Routines, Advisor Strategy, Claude Managed Agents) are tuned for Anthropic's models and may produce lower-quality multi-agent decomposition, but the core functionality ports cleanly.

Pricing and Access

Qwen3.7-Max is available through Alibaba Cloud Model Studio, with API access via both OpenAI-compatible and Anthropic-compatible endpoints. The model ID is qwen3.7-max. It is currently priced at $1.25 per million input tokens and $3.75 per million output tokens on some platforms, with a 50% promotional discount active at launch. Other sources list the standard rate at $2.50/$7.50 per million tokens. Cached input drops to $0.25 per million , a 90% discount that significantly changes the math for agentic workflows that reuse system prompts across hundreds of turns.

The model is also available on OpenRouter and Together AI. Together AI confirms the 1M token context window.

For context: Claude Opus 4.7 is $15/$75 per million tokens. GPT-5.5 is $10/$30. Qwen3.7-Max is the cheapest of the three frontier-tier models by a significant margin.

What It Cannot Do (Yet)

Three limitations are worth flagging before committing a production workload:

• Text-only. Qwen3.7-Max is text-only , text input, text output. For vision input, Alibaba provides Qwen3.7-Plus-Preview, which ranked #16 on LM Arena's Vision Arena leaderboard.
• No open weights. The model is API-only, routed through Alibaba Cloud's Singapore region. If your data residency policy requires US- or EU-only routing, this is a blocker. An open-weight variant is likely but not yet confirmed.
• Verbosity. Long-context retrieval reliability has not yet been independently verified across the full 1M window, so treat the ceiling as a capacity limit rather than a performance guarantee. Independent evaluators have also noted the model generates significantly more tokens than comparable models, which can inflate effective costs for long agentic sessions.

The Bigger Picture

Qwen3.7-Max arrives at a moment when the agentic AI market is genuinely competitive for the first time. The race for the best agentic AI model is getting crowded. Anthropic's Claude Opus 4.6, DeepSeek V4 Pro, and Kimi K2.6 have all staked claims to the top of the coding and reasoning leaderboards in recent months.

What Alibaba has done with Qwen3.7-Max is demonstrate that a Chinese lab can now compete at the frontier of agentic AI , not just on benchmarks, but on the specific capabilities that matter for production: long-horizon execution, tool orchestration, and cross-framework generalization. The price point makes the comparison even more uncomfortable for Western competitors.

For teams building coding agents, workflow automation, or any system that fires hundreds of tool calls per task, Qwen3.7-Max is now the rational default to evaluate. The cost difference compounds: a workflow that costs $300 in Opus 4.7 tokens costs roughly $50 in Qwen3.7-Max tokens , and on Terminal-Bench and SWE-Pro, Qwen leads. The only reasons not to switch are data residency requirements, dependence on Anthropic-specific agent features, or workloads where raw chat quality and enterprise document reasoning matter more than agentic coding performance.