CONFIRMED: Qwen ships the Max 0902 snapshot

Editorial label: CONFIRMED. On September 2, 2026 the official @Alibaba_Qwen account announced Qwen3.8-Max-0902, and the QwenCloud product page lists the model with alias qwen3.8-max-2026-09-02. This is not a new family: it is a post-trained checkpoint of the August Qwen3.8-Max, aimed at coding and cowork. Primary sources: https://x.com/Alibaba_Qwen/status/2094968708288680276 and https://www.qwencloud.com/models/qwen3.8-max-0902

What we know

The QwenCloud card describes gains on engineering-scale projects, long-horizon autonomous development, multi-tool orchestration and native vision (charts, documents, perception). It keeps a 1M context window, thinking mode and the full tool stack. Cloud pricing: $2 input / $6 output per million tokens; implicit cache $0.25; explicit cache read $0.17. Listed limits include 131K output, 262K reasoning, 1M TPM and 15K RPM. Qwen says 0902 scored 1,691 on CodeArena WebDev and took first place, 22 points above the prior Max.

What we still do not know

There is no long new qwen.ai blog just for 0902: this cycle's official record is the account post plus the API card. The 1,691 score and first-place claim come from Qwen citing CodeArena; this desk did not independently read the Arena.ai leaderboard in this cycle. Portuguese quality, SWE-bench Pro and real cost on long agent runs still lack a third-party harness for this snapshot. The August open weights of Max are not this checkpoint.

Why it matters

Teams already on Qwen3.8-Max can point the API at qwen3.8-max-0902 without changing the base rate. It is the cheap Grok 4.6-class $2 / $6 lane trying to take WebDev and cowork while Fable 5.1 still charges $10 / $50. Run an A/B on front-end apps and tool orchestration; do not migrate production on a ranking tweet. Gemini 3.8 Flash remains a rumor: the official Gemini API changelog still does not list the model.