Alibaba's Qwen3.8-Max codes autonomously for sixteen days — and undercuts Opus 5 by 60%
Qwen3.8-Max activates 95 billion of its 2.4 trillion parameters, holds a million tokens of context, and reportedly ran a sixteen-day unattended software project — reproducing an ML paper across 33 GPU training rounds, then inventing 18 improvements that beat it. Open weights are promised, and input pricing sits near 40% of Claude Opus 5.
The number that reframes the frontier is not the parameter count — it is sixteen days. A model that plans, writes, tests, reads its own logs, and iterates for two weeks without a human in the loop is not an assistant with a longer context window; it is an autonomous worker with a duty cycle. Every prior capability claim has been measured in a single response. This one is measured in calendar time.
The research reproduction is the harder evidence. Recreating a machine-learning paper from zero — 33 rounds of GPU training, roughly 125 hours, 7,600 lines of code — is a task with an objective pass mark, and then proposing 18 modifications that outperform the original crosses from replication into contribution. That is the exact motion the labs have promised for two years and rarely demonstrated on the record.
The pricing is the strategic blow. At roughly 40% of Opus 5's input cost and 24% of its output cost, with open weights following and a 27B variant behind it, Alibaba is not selling a competitor to the Western frontier — it is selling the same capability class as a commodity. The question the closed labs now have to answer is what their premium buys when the sixteen-day run is available at a discount and downloadable next week.
Developer Tech — Alibaba Qwen3.8-Max claims 16-day autonomous coding run → · Dataconomy — Alibaba unveils open-source Qwen3.8-Max AI model → · CGTN — Alibaba unveils Qwen3.8-Max, its most capable AI model to date →