Qwen3.8-Max: 2.4T coding flagship, $2/$6, open weights next week
Alibaba's flagship undercuts Opus 5 and GPT-5.6 on output price, tops Terminal-Bench, and charges one flat rate no matter how long your prompt is.
Copy markdown
2.4T MoE, and coding is where it jumps
Qwen3.8-Max is a 2.4-trillion-parameter mixture-of-experts model taking text, image and video in. DeepSWE 1.1 leaps from 21.6 to 56.6 over the last Max and Terminal-Bench 2.1 hits 86.6 — the gains land squarely in agentic coding.
$2 in, $6 out — it undercuts the frontier
At $2/$6 per million tokens it's a fraction of Claude Opus 5's $5/$25 and cheaper on output than GPT-5.6 Terra's $2/$12. Cached input drops to about $0.20/M, so a stable prompt prefix is where your bill actually shrinks.
One flat rate at any context length
The 1M-token window carries no long-prompt surcharge: a 5K prompt and a 900K prompt both bill at the base $2 input rate, unlike most frontier models. Feed it a whole repo without watching the meter climb.
OpenAI-compatible API, live today
The hosted endpoint ships now with a drop-in OpenAI-compatible schema, so pointing an existing client at it is a base-URL swap. Max input is 991K tokens (983K with thinking on), output up to 131K.
Open weights next week — including a self-hostable 27B
Alibaba says weights for both Qwen3.8-Max and a Qwen3.8-27B land next week, with the 27B pitched as the realistic on-prem path on standard GPUs. If it holds, you get a frontier-class coding checkpoint you own.
It tops the computer-use rows too
Beyond code it leads most vision benchmarks and scores 86.1 on OSWorld-Verified, the computer-use test — relevant if you're building agents that navigate real screens rather than just call APIs.