AlibabaProprietary

Qwen3.7-Max-Preview

Name: Qwen3.7-Max-Preview
Author: Alibaba

The latest agent-specialized model developed by Alibaba. It enables advanced agent workflows and reasoning capabilities.

Share:X Hatena LINE

Parameters

Undisclosed

Context Window

License

Proprietary

Release Date

2026-05-20

Japanese Language Capability

✅High-Quality JP

Multilingual model with strong Japanese language processing capabilities.

API Pricing

Input Price (per 1M tokens)

$2.5

Output Price (per 1M tokens)

Billing Mode: standard

Strengths

Weaknesses

Use Cases

Deep Analysis

Artificial Analysis Intelligence Index

56.6

#5 overall out of 218+ ranked models; highest Chinese model; +4.8 over Qwen3.6 Max Preview

Arena AI Elo

~1,489 (#14 overall)

#7 Math, #9 Expert Prompts, #9 Software & IT; WebDev Arena #4 with 1,541 Elo

SWE-Bench Verified

80.4%

SWE-Bench Pro: 60.6%; SWE Multilingual: 78.3%; leading domestic models

GPQA Diamond

92.4%

PhD-level reasoning; above Claude Opus 4.6 (91.3%), below GPT-5.5 (93.6%)

Input / Output Price

$2.50 / $7.50 per 1M tokens

Cached input: $0.25/1M (90% discount); ~half of Claude Opus 4.7 pricing

Context Window

1M tokens

4× increase from 256K on Qwen3.6 Max Preview; max output 64K–66K tokens

Strengths

・Best price-to-intelligence ratio among frontier models — matches or beats Claude Opus 4.6 on multiple benchmarks at roughly one-third the output cost
・Exceptional long-horizon agentic capability with 1M-token context and demonstrated 35-hour autonomous execution with 1,000+ tool calls
・Industry-leading abstinence behavior: lowest hallucination rate (22.9%) among frontier models on AA-Omniscience, with strong GPQA Diamond (92.4%) and HMMT (97.1%) scores

Weaknesses

・Text-only with zero vision capability — completely disqualifies any workflow requiring image understanding, UI screenshots, or computer-use agents
・High verbosity in reasoning mode generates ~97M output tokens vs. 26M median on benchmarks, significantly increasing latency (10–45s thinking overhead) and per-task API costs
・No open weights available; proprietary API-only access limits local deployment, fine-tuning, and offline use cases compared to earlier Qwen releases

Competitor Comparison

Model	Arena	SWE	GPQA	Price
GPT-5.5 (OpenAI)	~1,550+	N/A (published)	93.6%	$5.00/$30.00
Claude Opus 4.7 (Anthropic)	Top 5	64.3% (Pro)	~91–92%	$5.00/$25.00
Gemini 3.1 Pro Preview (Google)	Top 10	N/A	N/A	N/A (preview)
Gemini 3.5 Flash (Google)	Top 10	N/A	N/A	$1.50/$9.00
DeepSeek V4 Pro	Below Qwen3.7-Max	N/A	N/A	$1.74/$3.48

Overview

Qwen3.7-Max-Preview is Alibaba's flagship proprietary reasoning model, released on May 20–21, 2026, at the Alibaba Cloud Summit in Hangzhou. Positioned as an 'agent foundation' rather than a general-purpose chat assistant, it targets long-horizon autonomous coding, scientific reasoning, and office workflow automation. The model scored 56.6 on the Artificial Analysis Intelligence Index (#5 overall), making it the highest-ranked Chinese model on that leaderboard. It holds a 1M-token context window (4× the previous generation), delivers 92.4% on GPQA Diamond, 80.4% on SWE-Bench Verified, and demonstrated a 35-hour autonomous coding run on Alibaba's custom Zhenwu M890 accelerator in internal testing. The model occupies a distinctive sweet spot in the market: it approaches the raw intelligence of Claude Opus 4.7 and Gemini 3.1 Pro Preview while costing roughly half on input tokens and a third on output tokens. Its aggressive prompt caching ($0.25/1M cached input — a 90% discount) makes it particularly attractive for agent workflows that re-read long contexts across hundreds of turns. The model natively supports both OpenAI and Anthropic API specifications, enabling drop-in integration with tools like Claude Code. However, it is text-only (no vision), generates very verbose reasoning traces that inflate token costs, and has a notably high abstention rate (48% attempt rate on AA-Omniscience), which reduces hallucinations but limits usefulness on knowledge-recall tasks. Alibaba's release marks a strategic shift toward closed-weight, revenue-generating models at the top tier while maintaining open weights for lower-tier variants (announced 27B dense and 35B MoE models have no confirmed release date). The one-month cadence from Qwen3.6 to Qwen3.7 signals an aggressive development pace. For engineering teams, Qwen3.7-Max represents the strongest cost-optimized frontier option for text-only agentic and coding workloads, though teams requiring vision, absolute peak intelligence, or low-latency chat should look elsewhere.

Benchmarks & Performance

## Comprehensive Benchmark Performance ### Overall Intelligence | Benchmark | Qwen3.7-Max | GPT-5.5 | Claude Opus 4.7 | Gemini 3.1 Pro | Gemini 3.5 Flash | |---|---|---|---|---|---| | AA Intelligence Index | **56.6 (#5)** | 60.2 (#1) | 57.3 (#3) | 57.2 (#4) | 55.3 (#7) | | Arena AI Elo (Overall) | ~1,489 (#14) | Top 5 | Top 5 | Top 10 | Top 10 | | WebDev Arena Elo | 1,541 (#4) | — | — | — | — | ### Coding & Software Engineering | Benchmark | Qwen3.7-Max | Notes | |---|---|---| | SWE-Bench Verified | **80.4%** | Leading domestic models | | SWE-Bench Pro | **60.6%** | vs. Qwen3.6 Max: 57.3% | | SWE Multilingual | **78.3%** | Top position reported | | Terminal-Bench 2.0 | **69.7%** | vs. Qwen3.6 Max: 65.4%; one of highest public scores | | Terminal-Bench Hard | **50.8%** | +6.9pp over Qwen3.6 Max | | SciCode | **53.5%** | vs. Qwen3.6 Max: 47% | | LiveCodeBench | **91.6%** | — | | NL2Repo | **47.2%** | vs. Qwen3.6 Max: 42.9% | | AA Coding Index | **66.0%** | — | ### Scientific & PhD-Level Reasoning | Benchmark | Qwen3.7-Max | vs. Competitors | |---|---|---| | GPQA Diamond | **92.4%** | Above Claude Opus 4.6 (91.3%); below GPT-5.5 (93.6%) | | Humanity's Last Exam (HLE) | **41.4%** | vs. Qwen3.6 Max: 28.9%; Claude Opus 4.6: 40.0% | | CritPt | **13.4%** | +9.7pp over Qwen3.6 Max (3.7%) | ### Mathematics | Benchmark | Qwen3.7-Max | Notes | |---|---|---| | HMMT February 2026 | **97.1%** | Highest in comparison group | | IMOAnswerBench | **90.0%** | — | | Apex | **44.5%** | vs. DeepSeek V4 Pro: 38.3% | ### Agentic Capabilities | Benchmark | Qwen3.7-Max | Notes | |---|---|---| | Terminal-Bench 2.0-Terminus | **69.7%** | Above DeepSeek-v4-pro-Max and Claude Opus 4.6 | | QwenClawBench | **64.3%** | vs. Qwen3.6 Max: 59.0% | | GDPval-AA | **1,546 Elo** | +42 Elo over Qwen3.6 Max (1,504) | | Kernel Bench GPU | **1.98× acceleration** | Internal Alibaba testing only | ### Knowledge & Factual Recall | Benchmark | Qwen3.7-Max | Notes | |---|---|---| | AA-Omniscience Index | **14.1%** | 6th overall among frontier models | | AA-Omniscience Accuracy | **30.1%** | Dropped from 37.7% (Qwen3.6 Max) | | AA-Omniscience Hallucination Rate | **22.9%** | Lowest among frontier models; down from 44.2% | | AA-Omniscience Attempt Rate | **48.0%** | Lowest among frontier models — abstains on >half of prompts | | SuperGPQA | **73.6%** | vs. Qwen3.6 Max: 73.9% | | MMLU-Pro | **89.6%** | — | ### Instruction Following | Benchmark | Qwen3.7-Max | Notes | |---|---|---| | AA-IFBench | **80.5%** | vs. Qwen3.6 Max: 76.6% | | IFEval | **94.3%** | — | ### Speed & Efficiency | Metric | Qwen3.7-Max | Notes | |---|---|---| | Output Speed | **208 tokens/sec** | Tied for 3rd with Gemini 3.5 Flash | | Output Tokens on AA Eval | **~97M** | vs. median 26M — very verbose reasoning traces | ### Key Observation The model's standout characteristic is its extreme abstention strategy: it achieves the lowest hallucination rate among frontier models (22.9%) partly by refusing to answer more than half of knowledge prompts (48% attempt rate). This is a meaningful trade-off — excellent for tasks requiring factual reliability but limiting for broad knowledge retrieval.

Detailed Comparison

## Head-to-Head Comparisons ### Qwen3.7-Max vs. Claude Opus 4.7 | Dimension | Qwen3.7-Max | Claude Opus 4.7 | |---|---|---| | AA Intelligence Index | 56.6 (#5) | 57.3 (#3) | | GPQA Diamond | 92.4% | ~91–92% | | Context Window | 1M tokens | 1M tokens | | Input Price | $2.50/1M | $5.00/1M | | Output Price | $7.50/1M | $25.00/1M | | Cached Input | $0.25/1M | Available | | Vision | ❌ Text-only | ✅ Multimodal | | Open Weights | ❌ Proprietary | ❌ Proprietary | | API Compat | OpenAI + Anthropic | Anthropic native | **Verdict:** Qwen3.7-Max trails Opus 4.7 by less than 1 point on the Intelligence Index while costing one-third on output tokens. For pure text/code agentic workloads at scale, the pricing advantage is compelling. However, Opus 4.7 remains ahead on Arena rankings and supports vision. Choose Qwen3.7-Max for cost-sensitive text-only agent pipelines; choose Opus 4.7 when vision or the highest absolute quality ceiling is required. ### Qwen3.7-Max vs. GPT-5.5 | Dimension | Qwen3.7-Max | GPT-5.5 | |---|---|---| | AA Intelligence Index | 56.6 | 60.2 | | GPQA Diamond | 92.4% | 93.6% | | Context Window | 1M tokens | 1M tokens | | Input Price | $2.50/1M | $5.00/1M | | Output Price | $7.50/1M | $30.00/1M | | Vision | ❌ | ✅ Multimodal | **Verdict:** GPT-5.5 leads by 3.6 points on the Intelligence Index — the largest gap among the frontier tier. However, it costs 4× more on output tokens. For teams where per-task cost is the primary constraint and the 3.6-point intelligence gap doesn't materially affect task completion, Qwen3.7-Max offers dramatically better economics. ### Qwen3.7-Max vs. Gemini 3.5 Flash | Dimension | Qwen3.7-Max | Gemini 3.5 Flash | |---|---|---| | AA Intelligence Index | 56.6 | 55.3 | | Input Price | $2.50/1M | $1.50/1M | | Output Price | $7.50/1M | $9.00/1M | | Output Speed | 208 tok/s | 208 tok/s | | Vision | ❌ | ✅ | **Verdict:** Gemini 3.5 Flash is cheaper on input tokens and supports vision, making it better for high-volume, lower-stakes workloads. Qwen3.7-Max leads by 1.3 points on the Intelligence Index and is cheaper on output tokens, making it the better choice for reasoning-heavy agentic tasks where output token volume dominates cost.

Community Feedback

## Community & Industry Reaction **Positive Reception:** - LogRocket's June 2026 power rankings called Qwen3.7-Max "the month's biggest surprise" — the most-discussed model among five new frontier entries that month - The WebDev Arena #4 ranking (1,541 Elo, ahead of Claude Opus 4.6) generated significant buzz among developer communities, particularly for the price-to-performance ratio - Developers on forums and social media have highlighted the $0.25/1M cached input pricing as a game-changer for long-context agent workflows, with some reporting 70–90% cost reductions vs. Claude Opus for repetitive agentic tasks - The native Anthropic API compatibility (allowing drop-in use with Claude Code) was widely praised as a strategic move that lowered switching costs **Concerns & Skepticism:** - The 35-hour autonomous execution claim (1,158 tool calls) has not been independently verified — community consensus treats it as directionally plausible but unconfirmed - Multiple reviewers flagged the high abstention rate (48% attempt rate on AA-Omniscience) as a significant practical limitation: the model's low hallucination score is partly an artifact of refusing to answer, not superior knowledge - The shift from open to closed weights for Max-tier models drew criticism from the open-source AI community, with some noting it follows leadership turnover in the Qwen team - DeepLearning.AI's editorial noted being "saddened by Alibaba's turn toward closed weights" while acknowledging the quality of the model itself **Adoption Patterns:** - Early adoption is concentrated in Chinese enterprise teams already on Alibaba Cloud infrastructure and cost-sensitive Western developers building long-running agent pipelines - The model has gained traction in the coding-agent ecosystem (Cursor, Claude Code alternatives) due to strong SWE-Bench scores and aggressive pricing - Version pinning concerns: the monthly release cadence (3.6 → 3.7 in one month) has prompted community advice to pin specific model versions in production routing rather than using floating aliases like `qwen-max` - OpenRouter and Together AI adoption reported alongside direct Alibaba Cloud Model Studio usage

Use Cases

## Specific Use Cases & When to Choose Qwen3.7-Max ### 1. Long-Horizon Autonomous Code Migration **Example:** Migrating a large Python 2.x codebase to Python 3.12, or refactoring a monolithic service into microservices across thousands of files. **Why Qwen3.7-Max:** The 1M-token context window can hold an entire mid-sized repository in a single request. The model's demonstrated ability to sustain 1,000+ tool calls over 35 hours makes it suitable for overnight autonomous refactoring jobs. At $0.25/1M cached input tokens, re-reading the same codebase context across hundreds of agent steps is dramatically cheaper than competing frontier models. **When to choose over alternatives:** When Claude Opus 4.7's $25/1M output pricing makes multi-hour autonomous coding runs cost-prohibitive. When the codebase is text-only (no visual assets, no UI screenshots needed). ### 2. Multi-Step Scientific Research Synthesis **Example:** Processing and cross-referencing 50+ research papers, extracting methodology comparisons, identifying contradictions, and generating a structured literature review. **Why Qwen3.7-Max:** The 92.4% GPQA Diamond score demonstrates strong PhD-level reasoning. The 1M-token context can hold dozens of full papers simultaneously. The model's high abstention rate (48%) means it will flag uncertainty rather than hallucinate citations — a critical property for scientific work. **When to choose over alternatives:** When accuracy and citation reliability matter more than breadth of recall. Note: the model's 30.1% raw accuracy on AA-Omniscience means it may miss some retrievable facts — test against your specific literature corpus. ### 3. Office Workflow Automation & Document Processing **Example:** Automated contract review pipelines that process hundreds of legal documents daily, flagging clause variations, risk indicators, and compliance issues. **Why Qwen3.7-Max:** Alibaba specifically positions the model for office automation. The text-only limitation is less relevant here (contracts are text). The $2.50/1M input pricing with $0.25 caching makes high-volume document processing economically viable. The low hallucination rate is essential for legal/compliance contexts. **When to choose over alternatives:** When documents are text-only. If scanned PDFs or image-heavy contracts require OCR + vision understanding, use Qwen3.7-Plus-Preview or Gemini 3.5 Flash instead. ### 4. Cost-Optimized Coding Agent for Startups & Small Teams **Example:** A development team running a Cursor-like coding assistant that needs frontier-quality code generation at sustainable API costs. **Why Qwen3.7-Max:** At $2.50/$7.50 input/output — roughly half of Claude Opus 4.7's pricing — and 80.4% on SWE-Bench Verified (competitive with Opus 4.6), it delivers near-frontier coding quality at substantially lower cost. The native Anthropic API compatibility means it works as a drop-in for Claude Code-based agent harnesses. **When to choose over alternatives:** When the team is cost-sensitive and workflow is text/code-only. When the 1–2 point quality gap vs. Opus 4.7 doesn't affect task completion rates. Choose Opus 4.7 or GPT-5.5 when debugging requires visual inspection of UI or when absolute maximum code quality is non-negotiable.

Latest News

## Recent Developments (as of mid-2026) - **May 19, 2026:** Preview versions (Qwen3.7-Max-Preview and Qwen3.7-Plus-Preview) quietly appeared on Arena AI and Qwen Chat, ahead of formal announcement - **May 20–21, 2026:** Official announcement at the Alibaba Cloud Summit in Hangzhou. API access rolled out via Alibaba Cloud Model Studio (DashScope) - **May 2026:** API pricing confirmed at $2.50/$0.25/$7.50 per 1M input/cached/output tokens — notably, the free tier of Qwen Code was removed in the weeks prior - **May–June 2026:** Model made available on OpenRouter and Together AI for broader access - **June 1, 2026:** DashScope OpenAI Responses API support went live for Qwen3.7-Max on compatible-mode/v1, adding stateful previous_response_id, built-in web search and code interpreter, and reasoning.effort control - **Open-weight variants announced:** 27B dense and 35B MoE models announced but with no confirmed release date as of mid-2026. Historical pattern (Qwen3.6) suggests a ~6-point gap between Max and open-weight tiers - **Strategic shift confirmed:** Alibaba confirmed the Max tier will remain proprietary while the Plus tier will be open-source, continuing a pattern established with Qwen3.5. Leadership turnover in the Qwen team preceded this move - **Monthly cadence:** Qwen3.6 shipped April 2026; Qwen3.7 shipped May 2026 — indicating a 4–6 week release cycle for the Qwen series. Teams are advised to implement version-pinned routing

The model occupies a distinctive sweet spot in the market: it approaches the raw intelligence of Claude Opus 4.7 and Gemini 3.1 Pro Preview while costing roughly half on input tokens and a third on output tokens. Its aggressive prompt caching ($0.25/1M cached input — a 90% discount) makes it particularly attractive for agent workflows that re-read long contexts across hundreds of turns. The model natively supports both OpenAI and Anthropic API specifications, enabling drop-in integration with tools like Claude Code. However, it is text-only (no vision), generates very verbose reasoning traces that inflate token costs, and has a notably high abstention rate (48% attempt rate on AA-Omniscience), which reduces hallucinations but limits usefulness on knowledge-recall tasks.

Alibaba's release marks a strategic shift toward closed-weight, revenue-generating models at the top tier while maintaining open weights for lower-tier variants (announced 27B dense and 35B MoE models have no confirmed release date). The one-month cadence from Qwen3.6 to Qwen3.7 signals an aggressive development pace. For engineering teams, Qwen3.7-Max represents the strongest cost-optimized frontier option for text-only agentic and coding workloads, though teams requiring vision, absolute peak intelligence, or low-latency chat should look elsewhere.

Sources

Analysis generated: 2026-07-17