Qwen3.6‑Plus Nears Claude
Alibaba’s new closed-source Qwen3.6‑Plus matched Claude Opus 4.5 on SWE‑bench (56.6% vs 57.1%) and beat it on Terminal‑Bench (61.6% vs 59.3), signaling stronger coding and agentic task performance from an Asian model vendor. (x.com)
Alibaba published an official blog post for Qwen3.6‑Plus on April 2, 2026 announcing the model and its enterprise availability via Alibaba Cloud Model Studio. (alibabacloud.com) A public preview of Qwen3.6‑Plus appeared on OpenRouter on March 30–31, 2026 under the model ID qwen/qwen3.6-plus-preview with free limited access. (openrouter.ai) Qwen3.6‑Plus ships with a 1,000,000‑token context window and explicit reasoning/thinking modes designed for tool use and long‑horizon agentic workflows. (alibabacloud.com) Early API evaluations reported a SWE‑bench Verified result of 78.8% and a Terminal‑Bench 2.0 score of 61.6% in third‑party test runs. (apidog.com) The model exposes a preserve_thinking parameter (disabled by default) that controls whether prior "thinking" tokens are retained across conversation turns to support multi‑step agent loops. (alibabacloud.com) Bloomberg and other outlets framed Qwen3.6‑Plus as one of several recent closed‑source/proprietary releases from Alibaba as the company pivots toward monetizing flagship AI models. (bloomberg.com) OpenRouter usage metrics published during the preview showed roughly 400 million completion tokens served across about 400,000 requests in the model’s first two days. (apidog.com) Terminal‑Bench leaderboards continue to feature agentic setups built on Anthropic’s Claude Opus 4.6 and OpenAI’s GPT‑5 family at the top, indicating Qwen3.6‑Plus is entering an active comparative field for agent and coding benchmarks. (tbench.ai)
