QWEN3.6-MAX
Closed-weight flagship released April 20, 2026. #1 on SWE-bench Pro, Terminal-Bench 2.0, SciCode and three more. 262K context. Native web search. $1.04/$6.24 per million tokens.
Qwen3.6-Max is Alibaba's closed-weight flagship, released April 20, 2026. It currently holds the #1 spot on six top coding benchmarks — SWE-bench Pro, Terminal-Bench 2.0, SciCode, and three more — beating GPT-5.5, Claude Opus 4.7, and Gemini 3 Pro on agentic programming evaluations as of release. The context window is 262K tokens, and native web search via DashScope is enabled. Priced at $1.04 per million input tokens and $6.24 per million output, it sits between DeepSeek V4 ($0.14/$0.28) and GPT-5.5 ($5/$30) — about 5x cheaper than GPT-5.5 on input and roughly 5x cheaper on output. The case for: serious coding work, agentic dev loops, SWE-bench-style refactors at a fraction of frontier US-model pricing. The case against: closed weights (no self-hosting), Anthropic and OpenAI still have stronger ecosystems for tool calling in production agents, and Opus 4.7 still wins on writing voice and editorial nuance.
Qwen3.6-Max is the newer (April 2026) closed-weight flagship with stronger coding benchmark results. Qwen3-Max is the prior trillion-parameter flagship with thinking mode.
No. Unlike many earlier Qwen releases, the Max tier is closed weights. You access it only through Alibaba's DashScope API (or via Council AI).
As of April 2026 release: SWE-bench Pro, Terminal-Bench 2.0, SciCode, and three others — six top coding benchmarks in total.
$1.04 per million input tokens and $6.24 per million output. About 5x cheaper than GPT-5.5.
On benchmark numbers, yes — Qwen3.6-Max edges Opus on SWE-bench Pro and Terminal-Bench. In practice, run both in a Council AI council and compare outputs on your actual codebase.
Yes — natively via DashScope's enable_search flag. It can ground answers on live web results.