Qwen3.8-Max's coding checkpoint hits OpenRouter at $2/$6, 1M context
Alibaba's 0902 snapshot doubles its terminal-agent scores and edges Opus 5 Max on web dev — a 2.4T flagship at a fifth of Astra's price, one weak spot aside.
Copy markdown
A cheap, huge-context coding model on OpenRouter
The Sept 2 '0902' checkpoint is live on OpenRouter at $2/M input and $6/M output, with a 1M-token context and 131K max output — roughly a fifth of GPT-6 Astra's $10/$50. Alibaba Cloud serves it directly, and it's OpenAI-compatible, so pointing an existing agent at it is a base-URL swap.
The coding tune more than doubled its agent scores
Against the base model, TerminalBench 3.0 jumped from 11.3 to 29.0 and ProgramBench Almost-Solved from 10.5 to 28.0. On Code Arena WebDev it posts 1,691 — three points above Claude Opus 5 Max — and it tops OSWorld computer-use at 86.1, ahead of Fable 5, GPT-5.6 Sol Max, and Gemini 3.1 Pro.
But SWE-bench Pro is its soft spot
On SWE-bench Pro it lands at 67.7 — behind Fable 5 (80.0) and Opus 4.8 (69.2), though ahead of GPT-5.6 Sol (64.6). Read that as: strong on terminal loops, computer use, and web dev; weaker on gnarly real-repo bug-fixing. Route hard debugging elsewhere and use this for the rest.
2.4T MoE, multimodal in, tools and structured output
It's a 2.4-trillion-parameter mixture-of-experts model that accepts text, image, and video and returns text, with native tool calling, structured outputs, and a configurable reasoning dial. Throughput on OpenRouter runs ~36 tok/s at ~2.2s to first token — steady rather than blazing.
Your agent may already be routing to it
OpenRouter lists Claude Code, Portkey, and Hermes Agent among the model's top traffic sources — builders are already sending agent work here for the price and context. If you run long-horizon or multi-tool jobs, it's worth A/B-ing against your current default before the next bill.