Alibaba Launches Qwen3.8-Max: 2.4 Trillion Parameters and 60% Cheaper Than Claude Opus 5

On August 3, 2026, Alibaba's AI research division officially released Qwen3.8-Max, its largest and most capable model to date. At 2.4 trillion total parameters — activated in 95-billion-at-a-time chunks via a sparse Mixture-of-Experts architecture — it is one of the biggest models anyone has publicly deployed.

The number is large, but the architecture is what makes the model actually run. The MoE design means that on any given token, only a fraction of the network is live, keeping inference costs manageable despite the enormous parameter count. Combined with a hybrid attention mechanism, Qwen3.8-Max supports contexts up to 1 million tokens, putting it in direct competition with the best long-context offerings from Anthropic and Google.

What it can do

Qwen3.8-Max accepts text, images, and video as input, returning text. In Alibaba's own evaluation, the model completed a multi-day software engineering project from scratch without human intervention, landing fourth on the Frontend Code Arena leaderboard with a score of 1,668. It ranks fifth overall in Text Arena and second in Vision Arena — positions that place it firmly in the same tier as Anthropic's Claude Opus 5 and Google's Gemini 3.6.

Alibaba highlighted its performance on long-horizon tasks: the kind of work requiring planning across dozens of steps rather than single-pass responses. Whether that holds up under third-party benchmarking remains to be seen, but the self-reported numbers are competitive across both reasoning and multimodal dimensions.

The pricing angle

The most interesting detail for developers is the price. Alibaba opened API access on Alibaba Cloud Model Studio at roughly 40% of Claude Opus 5's input-token cost and 24% of its output-token cost in international markets. At those numbers, workloads that were economically marginal with Western flagship models become viable at scale.

Model weights are scheduled for public release in the coming days, enabling self-hosted deployment — a significant option for teams with data-residency requirements or the hardware to run inference locally.

The broader context

Qwen3.8-Max continues a pattern that has defined Chinese AI labs in 2026: releasing capable models at prices that undercut Western competitors by a wide margin. Alibaba's previous Qwen 3.7 Flash already set a benchmark for low-cost long-context processing in early 2026. The new flagship pushes that cost pressure up the capability ladder.

For anyone evaluating frontier models, Qwen3.8-Max is now a credible alternative to Claude Opus 5 and GPT-5.6 for production workloads — provided Alibaba's benchmarks hold up to independent scrutiny. The open-weight release next week will give the research community a chance to find out.