Qwen3.8-Max ships: a 2.4T model you can wire into Claude Code
2.4T params, 95B active, near-1M context. It trails GPT-5.6 Sol and Fable 5 on coding, but runs $2/$6 per MTok — and open weights ship next week.
Copy markdown
The headline: 2.4T total, 95B active, Qwen's most capable yet
Qwen3.8-Max is a mixture-of-experts model with 2.4 trillion total parameters and ~95B active per token, now generally available on a hosted API. It hits 92.6 on GPQA Diamond and 86.1 on OSWorld-Verified — Alibaba's clearest gains this generation are agentic and multimodal, not raw reasoning.
Coding: a big generational jump, still a notch behind the frontier
It scores 67.7 on SWE-bench Pro (vs Fable 5's 80.0) and 86.6 on Terminal-Bench 2.1 (vs GPT-5.6 Sol's 88.8). The leap is the story: FrontierSWE went 40.7 to 73.5 and DeepSWE 21.6 to 56.6 — good enough to be your default cheap coder, not your hardest-task model.
Wire it into Claude Code or Codex tonight
QwenCloud exposes both OpenAI- and Anthropic-compatible endpoints, so you swap a base URL and model ID and keep your existing harness. Qwen ships setup guides for Claude Code, Codex, Qwen Code, Qoder and OpenClaw, and a reasoning_effort flag (low/medium/xhigh) trades cost for depth.
The price is the point: $2 in, $6 out
Pricing is $2.00 / $6.00 per million tokens with cached input at $0.25 — a fraction of frontier closed-model rates for near-frontier agentic scores. With ~991K input and 131K output tokens, it's a cheap driver for long-context repo work.
Open weights — and a self-hostable 27B — land next week
Alibaba says the full weights hit Hugging Face and ModelScope next week, alongside an open Qwen3.8-27B sized to run on a single box. If that ships, it's the weekend project: a private agent backend with no API bill.