Alibaba releases Qwen3.8-Max-0902 with 1M-token context

• Qwen3.8-Max-0902 is live on QwenCloud's API with a 2.4T-parameter headline and a 1M-token context window, according to Alibaba_Qwen's release post.
• The updated model reached 1,691 on Code Arena: WebDev, a 22-point move from the prior Qwen3.8-Max score of 1,669, as Alibaba_Qwen's leaderboard post reported.
• API list pricing remains $2 per million input tokens and $6 per million output tokens, while Alibaba_Qwen's release post lists explicit and implicit cache reads at $0.17 and $0.25 per million tokens.
• Alibaba separately says Qwen3.8-Max is the strongest open-weight model on its commerce-agent evaluation, in Alibaba_Qwen's CommerceAgentBench post, though that post supplies no score table or methodology.

Code Arena's live table marks the 1,691 score as preliminary and gives it a ±19 spread. The open-source E-Commerce Bench repository runs agents through a simulated 365-day retail year, while an API catalog analysis found the dated release name has not yet appeared consistently as a distinct model identifier.

Alibaba describes 0902 as an upgrade to Qwen3.8-Max, further post-trained on coding and “Cowork.” It claims gains on enterprise work, scientific research, and long-horizon workflows, plus refined vision for charts and documents in the linked QwenCloud release page.

The commercial details are unusually specific:

• Input: $2 per million tokens.
• Output: $6 per million tokens.
• Explicit cache hit: $0.17 per million tokens.
• Implicit cache hit: $0.25 per million tokens.
• Context: 1 million tokens.

The announcement gives the total parameter count, not an active-parameter count or an architecture description.

The independent Code Arena leaderboard lists Qwen3.8-Max-0902 first at 1,691, above Claude Opus 5 Max at 1,687 and Kimi K3 Max at 1,674. Its 1,390 votes are far fewer than Claude's 10,517, and the row labels the Qwen result preliminary with a +19/-19 score range.

Alibaba's own announcement calls the improvement 1,669 to 1,691, while Alibaba_Qwen's WebDev claim attributes the jump to multistep reasoning, tool use, and full-app generation. The published score intervals overlap with Claude's, so the rank is an early leaderboard result rather than a settled separation.

Arena's Pareto-frontier post also puts 0902 on the WebDev price-performance frontier at a stated blended $5 per million tokens. The same post says the prior Qwen3.8-Max, Claude Opus 5 Max, and Kimi K3 Max sat at overall ranks two through four but fell off that frontier; the live table displays Qwen's API price as $2 input and $6 output.

Alibaba's CommerceAgentBench post says its benchmark begins with “real commercial demand” and that Qwen3.8-Max has the strongest overall open-weight result. The post does not identify the task set, scoring method, or exact model revision.

A different Qwen research project, E-Commerce Bench, supplies those mechanics: each agent starts with ¥100,000, operates up to four stores across 365 simulated days, and is scored on seven dimensions. The repository defines the primary result as end-of-year assets divided by the opening balance, averaged across five independent episodes.

The paper's reported open-weight leader is Qwen3.8-Max-Preview, not 0902: ¥416,252 in end-of-year assets, 38% above GLM 5.2 High, according to dair_ai's paper summary. It also reports no model led every dimension, separating asset accumulation from fraud avoidance, operational efficiency, negotiation, cash flow, execution, and long-horizon learning.

Qwen's public release label is Qwen3.8-Max-0902, but the API catalog analysis found Alibaba's documentation still naming the callable model qwen3.8-max. Alibaba_Qwen's release post says the update is available through QwenCloud, yet does not state whether -0902 is a separately pin-able API slug or a revision served behind the family name.