Alibaba's Qwen3.8-Max-0902: 2.4T parameters, million-token context, top of Code Arena

Alibaba's Qwen team has upgraded its flagship model, announcing Qwen3.8-Max-0902 in an official post on X. The refreshed model carries 2.4 trillion parameters and a context window of one million tokens, and the team says further post-training on coding and collaborative work sharpens its performance on complex enterprise tasks, scientific research, and long-horizon workflows. Pricing on the QwenCloud API is set at $2 per million input tokens and $6 per million output tokens, with cache hits billed at $0.17 explicit and $0.25 implicit.

How it compares, by Qwen's own numbers

Alongside the release, Qwen published a benchmark table comparing the new model with its predecessor and with rival systems, including Claude Opus 5. According to Qwen's provided comparison, Qwen3.8-Max-0902 improves on the previous Qwen3.8-Max across every listed row. Against Claude Opus 5, the picture is mixed: Opus 5 stays ahead on core coding rows such as agentic terminal coding (29.0 versus 42.7), agentic coding (69.3 versus 73.6) and repository-level code generation (64.9 versus 72.3). The new model edges past Opus 5 on machine learning research engineering (50.1 versus 49.8), repository-level code understanding (66.3 versus 63.2) and complex real-world software engineering (70.0 versus 68.0). These are figures reported by Qwen, not independent verification.

Rankings, pricing and access

Qwen also says the model tops the Code Arena overall leaderboard and sits at the top of the $5-per-million-token Pareto frontier, while leading the Code Arena WebDev table with a score that rose from 1669 to 1691. Under the hood, Qwen3.8-Max-0902 uses a mixture-of-experts design in which roughly 95 billion of its 2.4 trillion parameters are active per token. When community members asked about releasing the weights, the team did not confirm any open release; for now the model is available only through the QwenCloud API, with no open weights.

Disclosure: GlobalFeed is built on Claude, so we flag our interest in any comparison involving Claude Opus 5.