Qwen3.8-Max ships: 2.4T MoE, $2/$6, beats GPT-5.6 on computer use
Alibaba's biggest model is live on API at roughly half GPT-5.6's price, with 1M context, a 27B open-weights companion, and the Max weights due next week.
Copy markdown
2.4 trillion params, and a computer-use claim
Qwen3.8-Max is a sparse mixture-of-experts (~2.4T total, ~95B active per token) taking text, image and video. Alibaba says it beats GPT-5.6 Sol Max and Claude Fable 5 on agentic computer use, posting 86.1 on OSWorld-Verified.
Half the frontier's price
The API runs $2 per million input tokens and $6 output, with cached reads at $0.25 — down from the prior Max's $2.50/$7.50 and well under comparable GPT-5.6 and Opus tiers. Rate limits reach 2M tokens per minute.
Live now, OpenAI-compatible endpoint
The hosted API is broadly available today via DashScope and an OpenAI-compatible endpoint, so you can point an existing agent at it by swapping the base URL and key. Context is 1M tokens (991K in, 131K out).
Open weights: 27B companion, Max next week
A Qwen3.8-27B variant is going open-weights for on-prem GPUs, and Alibaba says the full Max weights land next week — unusual at this tier. The activated-parameter count for Max stays undisclosed, which muddies cost modeling.
Coding: strong, with an asterisk
It hits 86.6 on Terminal-Bench 2.1 (just behind GPT-5.6 Sol's 88.8, ahead of Opus 4.8's 84.6) but trails on SWE-bench Pro at 67.7 vs Claude Fable 5's 80.0. A huge jump over 3.7-Max, not a clean coding sweep.
Build this weekend: a cheap executor tier
With 1M context and sub-frontier pricing aimed at "coding and cowork," it's built for long agent runs where token cost dominates. Slot it in as the executor and keep a pricier model for the hardest reasoning.