Alibaba Releases Qwen3.8-Max, Performance Nears Claude, Price at 40% of Opus 5

Alibaba has surprise-launched its next-generation foundation model, Qwen3.8-Max, with a total of 2.4 trillion parameters. According to QbitAI, on the latest authoritative third-party leaderboard Arena, Qwen3.8-Max has broken into the global top tier, rivaling Anthropic's Claude series. In coding performance, Qwen3.8-Max even surpasses Claude Opus 5 and Fable 5's High version.

An extreme case disclosed by the company shows the model can start from an empty folder and autonomously advance for over ten days without human intervention, ultimately delivering a real, usable complete project. In long-horizon tasks, it demonstrates system-level autonomous planning and execution, staying on target across thousands of rounds of ultra-long interactions. For professional work, it can deliver an app prototype that would typically require 3 to 5 rounds of revisions in one go, and can process hundreds of legal documents in an hour, completing annotations for over a thousand clauses.

Qwen3.8-Max's pricing strategy is highly competitive: domestic pricing is 12 RMB per million tokens for input and 36 RMB for output, with implicit cache hits at only 1.5 RMB. According to QbitAI, international prices for input and output are only 40% and 24% of Opus 5's, respectively. This pricing makes high-performance models no longer expensive; netizen feedback indicates the price is reasonable, with weekly usage quotas not even 90% consumed.

QbitAI comments that model capability, scenario coverage, and price are often hard to achieve simultaneously, but Qwen3.8-Max attempts to break this deadlock. Behind it is Alibaba's cumulative open-sourcing of over 400 models since 2023, with more than 200,000 derivative models and cumulative downloads exceeding 1 billion, forming the world's largest and most comprehensive open-source model family.

In QbitAI's hands-on test, after receiving a full page of product requirements, Qwen3.8-Max delivered a near-production-grade full-stack solution in about 5 minutes, including requirement breakdown, technology selection, project structure, feature implementation, and automatically compiled problem and resolution logs. It even performed feature acceptance, checking each step from dependency installation to production build and marking them as passed.

In a long-text analysis test, Qwen3.8-Max was given a dozens-page paper (including main text, figures, and appendices) and asked to cross-analyze the difficulty of the ILSVRC and PASCAL VOC datasets. After about 33 seconds, the model gave a clear answer: looking at overall averages, PASCAL VOC is harder, but when considering comparable categories and difficult subsets, ILSVRC is not easy and even harder. QbitAI verified this against Table 3, Figure 16, and Appendix B, confirming the answer was correct.

In a multimodal content understanding test, Qwen3.8-Max was asked to process a nearly 70-minute video podcast (a Sam Altman interview) and generate a complete topic timeline, locating original segments by topic and viewpoint. In about two to three minutes, the model output content divided by core viewpoints, with each topic structured as a summary followed by elaboration, and corresponding timestamps annotated.

QbitAI commented that Qwen3.8-Max's work goes beyond generation; it can advance complete engineering projects from scratch and solve real problems. Currently, Qwen3.8's API is available on the Qianwen AI platform and has been integrated into Alibaba's concurrently launched Agent product, 'Qianwen Office'.