Qwen3.6-Plus

Overview

Qwen3.6-Plus is Alibaba's hosted-only flagship model for agentic coding, released April 2, 2026, after a free soft launch on OpenRouter two days earlier. It runs a 1-million-token context window by default, caps output at 65,536 tokens, and keeps chain-of-thought reasoning on for every single request. There's no toggle to switch it off.

That last detail is a real product decision, not a footnote. Alibaba built Qwen3.6-Plus around three pitches: agentic coding at the repository level, visual coding from UI screenshots and wireframes, and multimodal document reasoning. Our launch coverage traced the enterprise angle, where the model powers Alibaba's invitation-only Wukong platform and the consumer-facing Qwen App. This card is about whether the benchmark numbers back up the pitch.

Parameter count is undisclosed, as it has been for every Qwen flagship since the team stopped publishing MoE routing details. Third-party analysis pins the architecture as a hybrid combining linear attention with sparse mixture-of-experts routing, but Alibaba hasn't confirmed that description. What's verifiable is the behavior: fast long-context retrieval, reasoning turned on by default, and a pricing structure aimed squarely at high-volume agentic workloads rather than casual chat.

Key Specifications

Pricing varies by source and provider. Artificial Analysis and Together AI both list $0.50/$3.00 per million tokens as the standard rate. OpenRouter currently shows a promotional $0.325/$1.95, a 35% discount off the DashScope list price. Some trackers also cite a second Alibaba Cloud tier - $2.00/$6.00 per million tokens - for prompts between 256K and 1M tokens, though Alibaba hasn't published that split on its own pricing page. Treat the higher tier as unverified until Alibaba confirms it directly.

Benchmark Performance

Two figures for GPQA Diamond circulate in third-party coverage - 90.4% from BenchLM and Together AI, and a lower 88.2% from a separate aggregator. Alibaba hasn't published a GPQA number on its own blog post, so neither figure carries a primary-source stamp. The 90.4% reading has two independent corroborations against one for the alternative, which is why it's the number in the table above.

The pattern that holds up across sources: Qwen3.6-Plus wins decisively on document understanding (OmniDocBench, where it beats Kimi K2.5, GLM-5.1, and both Claude 4.5 Opus and Gemini 3 Pro) and knowledge-heavy academic benchmarks, but trails on the agentic coding benchmarks that matter most for its stated use case. Head-to-head against Kimi K2.6, our aggregated numbers put Kimi ahead on average agentic-task performance (73.1 vs 61.6) while Qwen leads on average knowledge-task performance (66.0 vs 53.8). Neither model dominates the other; they trade wins by category.

On our own site's vision-language benchmarks leaderboard, Qwen3.6-Plus posts 78.8% on MMMU-Pro, 85.4% on RealWorldQA, 81.5% on CharXiv-R, and a category-leading 94.4% on AI2D. The web agent benchmarks leaderboard has it at 57.2% on WebArena, seventh among tracked models and behind Gemini 3.1 Pro's 58.4%. On the instruction-following leaderboard, it scores 94.3% on IFEval, second only to the open-weight Qwen3.5-27B in that category.

Qwen3.6-Plus runs exclusively through Alibaba Cloud infrastructure - there's no self-hosting path since Alibaba never published weights for this tier. Source: commons.wikimedia.org

Key Capabilities

Agentic coding at repository scale. The 1M-token context window means Qwen3.6-Plus can hold an entire mid-sized codebase in a single session rather than chunking it across calls. Combined with the always-on reasoning, that's the intended workflow: plan across a full repo, execute a multi-step fix, and verify against a test suite without losing earlier context. The 61.6 Terminal-Bench 2.0 score edges out Claude 4.5 Opus's 59.3 on the same suite, though Kimi K2.6 clears both at 66.7.

Terminal-Bench 2.0 and SWE-bench Verified measure exactly this kind of multi-file, multi-step coding work rather than single-function completions. Source: commons.wikimedia.org

Visual coding from screenshots. Feed it a UI screenshot, a hand-drawn wireframe, or a Figma export, and it produces working frontend code. This isn't new in isolation, but the long context window changes what's feasible for design systems with hundreds of components - an entire component library can fit in one prompt where competitors need to paginate.

Document-heavy reasoning. The 91.2 OmniDocBench v1.5 score is the strongest result in the lineup, ahead of every model Alibaba benchmarked it against including Gemini 3 Pro and Claude 4.5 Opus. For teams parsing dense PDFs, scanned forms, or technical manuals, this is the category where Qwen3.6-Plus has the clearest lead over its named competitors.

The reasoning tax

Always-on chain-of-thought means every request pays a latency and token cost, even a one-line completion that a non-reasoning model would answer instantly. Cursor's agent tooling and similar products have shown that mixing reasoning and non-reasoning calls per task is usually cheaper than blanket reasoning. Qwen3.6-Plus doesn't give you that choice.

Pricing and Availability

Standard API access runs through Alibaba Cloud Model Studio (DashScope) at $0.50 per million input tokens and $3.00 per million output tokens, per Artificial Analysis and Together AI. OpenRouter currently lists a discounted $0.325/$1.95 - a 35% cut off list price, presumably promotional. There's no confirmed long-context surcharge from Alibaba directly, though some trackers report a $2.00/$6.00 tier above 256K tokens that hasn't been independently confirmed.

Against Kimi K2.6, which lists at $0.95/M input and $4.00/M output, Qwen3.6-Plus is meaningfully cheaper on both ends - roughly half the input cost and about a quarter less on output, at the DashScope standard rate. Against Claude Opus 4.7, the gap is larger still. Whether that price gap survives contact with the reasoning tax in practice depends on how verbose the model runs per task; Artificial Analysis's Intelligence Index evaluation logged roughly 100 million output tokens for Qwen3.6-Plus against a 62 million field average, a 61% verbosity premium that eats into the headline discount.

Speed claims disagree substantially across trackers, which is itself worth flagging. Artificial Analysis measured 53.4 tokens/second output and a 2.57-second time to first token, calling it "notably slow" relative to peers. A separate review from Digital Applied measured 158 tokens/second median throughput and roughly 11.5-second time to first token on the free preview tier, and framed it as 1.7x faster than Claude Opus 4.6. Both can't be describing the same production configuration; the gap likely comes down to provider routing and preview-tier rate limits rather than the model itself. Don't trust either number without testing your own workload.

There's no open-weight release for Qwen3.6-Plus. The nearest self-hostable sibling is Qwen3.6-35B-A3B, a 35B MoE with 3B active parameters, free under Apache 2.0, which scores 73.4% on SWE-bench Verified at a fraction of the cost. If self-hosting is a requirement, that's the model to look at instead.

Strengths

• Best-in-class document understanding: 91.2 on OmniDocBench v1.5, ahead of Claude 4.5 Opus, Gemini 3 Pro, and Kimi K2.5
• 1M native context window handles full repositories and large design systems in a single session
• Cheaper than Kimi K2.6 and Claude Opus 4.7 on both input and output pricing at the standard DashScope rate
• Terminal-Bench 2.0 score of 61.6 beats Claude 4.5 Opus's 59.3 on the same suite
• Strong instruction-following: 94.3% on IFEval, second only to Qwen3.5-27B among tracked models

Weaknesses

• No weights released - no self-hosting, no fine-tuning, no air-gapped deployment
• Trails Kimi K2.6 on the agentic benchmarks that matter most for its stated use case: SWE-bench Verified (78.8% vs 80.2%) and Terminal-Bench 2.0 (61.6 vs 66.7)
• Mandatory reasoning mode with no off-switch adds latency and cost to every request, including trivial ones
• Verbosity runs roughly 61% above the field average during independent evaluation, eating into the price advantage
• Speed benchmarks from independent trackers disagree by 3x on output tokens/second - treat any single speed claim skeptically
• A separate review flagged a 26.5% fabrication rate on reasoning claims about APIs and language behavior, a figure Alibaba hasn't addressed publicly

Related Coverage

• Alibaba's Qwen3.6-Plus Launches With 1M Context Window - our launch-day news coverage
• Qwen3.6-Max-Preview Model Card - the closed flagship that shipped three weeks later
• Our Qwen3.6-Max Review - hands-on benchmarking of the tier above Plus
• Qwen3.6-35B-A3B Model Card - the open-weight sibling for teams that need to self-host
• Kimi K2.6 Model Card - the open-weight competitor that beats Plus on most agentic benchmarks
• Coding Benchmarks Leaderboard - full SWE-Bench, Terminal-Bench, and LiveCodeBench rankings
• Vision-Language Benchmarks Leaderboard - where Qwen3.6-Plus ranks on MMMU-Pro, AI2D, and CharXiv-R
• Cost-Efficiency Leaderboard - pricing-to-performance comparisons across the field

Sources

• Qwen3.6-Plus: Towards Real World Agents - qwen.ai
• Qwen3.6 Plus Preview: 1M Context, Speed & Benchmarks - buildfastwithai.com
• Qwen3.6 Plus Benchmarks, Pricing & Speed - benchlm.ai
• Qwen3.6-Plus API - together.ai
• Qwen3.6 Plus - Intelligence, Performance & Price Analysis - artificialanalysis.ai
• Qwen 3.6 Plus: 1M Context With Always-On Reasoning - digitalapplied.com
• Kimi K2.6 vs Qwen 3.6 Plus: Two Chinese Frontier Models Compared - aimadetools.com
• Qwen3.6 Plus - API Pricing & Benchmarks - openrouter.ai
• Qwen 3.6 Plus - vals.ai
• Qwen3.6 Plus Benchmarks, Pricing & Context Window - llm-stats.com
• Alibaba press release via EQS News