Within a few weeks of each other, two AI advertisements landed. The first begins with a house engulfed in flames. The second could pass for the cover art on a lo-fi study playlist, all soft edges and calm, as a monitor churns through work and the people in frame head out to go rock climbing.
No benchmark chart captures the state of the AI race as neatly as that contrast does. The benchmarks still count for something, though, so we’ll begin with them.
Alibaba has now shipped its strongest model to date, handing every American tech executive already nervous about Chinese labs closing the gap one more thing to lose sleep over.
125 hours of unsupervised work
Qwen3.8-Max is the name. The model packs 2.4 trillion parameters and was designed for long-horizon agentic tasks — jobs that unfold over days “with minimal human involvement,” in the wording Alibaba used in its announcement.
One figure is hard to shake. During an internal test, Qwen3.8-Max operated by itself for something like 125 hours — nearly five days — while reproducing an experiment laid out in a research paper. The paper was all Alibaba gave it. Everything else the model handled alone: writing the code from scratch, crunching the data, running the experiment, then reporting the outcome.
According to Alibaba, that is “exactly the type of work that takes skilled engineers days.”
A five-day autonomous run is a very different sort of claim than a benchmark score. Extended agentic sessions are precisely where models tend to come undone, wandering away from the task or getting stuck in loops. So far Alibaba hasn’t released enough detail for anyone outside the company to judge how clean that particular run really was.
Where it lands against GPT-5.6 Sol and Fable 5
The test results in Alibaba’s announcement put Qwen3.8-Max shoulder to shoulder with OpenAI’s GPT-5.6 Sol and Anthropic’s Fable 5 across a range of benchmarks. On two of them — visual reasoning and agentic computer use — it comes out ahead of both.
Figures a lab publishes about its own model always warrant a raised eyebrow. The open weights are due out next week, so independent verification shouldn’t be far behind.
The price is the actual weapon
Anthropic charges $10 per million input tokens and $50 per million output tokens for Fable 5. Alibaba says Qwen3.8-Max will be priced at $2 per million input tokens and $6 per million output tokens.
Call it five times cheaper on the way in and better than eight times cheaper on the way out. For anyone running agentic workloads that chew through tokens for days on end, a gap that wide stops looking like a line item and becomes the entire budget discussion.
A good number of U.S. users have run those figures already. Many have been cancelling OpenAI and Anthropic subscriptions in favor of less expensive Chinese options.
Alibaba isn’t alone in the chase
Kimi K3 arrived last week from Chinese lab Moonshot. The test results it published likewise put the model just a step behind the most advanced American systems.
Anxiety about a Chinese lead has been building across Silicon Valley and inside the U.S. government for months now, as Chinese models — a great many of them open source — draw level with OpenAI, Anthropic and Google.
The ad that opens on a burning house
“There’s hope in hard questions,” Anthropic’s ninety-second spot, made its debut during last month’s World Cup. Early on it carries the feel of an A24 horror trailer — jarring notes, imagery that’s chaotic and unsettling — before easing into something resembling lukewarm hope.
The intent seems to be reassurance: proof that Anthropic understands the dangers and means to sidestep them, guiding humanity toward prosperity and well-being instead. Many viewers came away finding the ad tasteless, disturbing, or some combination of the two.
It opens on a burning house. Somewhere in it there is also what looks like a photograph of gravestone rows at Arlington National Cemetery — an ethically dubious image to drop into an ad for a product.
Friend, the American firm behind the much-debated AI pendant, mined similar territory last week. In its latest ad, two people share their personal struggles with the pendant while a soundtrack plays that could hardly be described as cheerful.
Alibaba’s answer: fishing, tennis, rock climbing
The commercial Alibaba put out alongside Qwen3.8-Max sits at the opposite end of the spectrum. It’s assembled like one of those bottomless YouTube loops labeled “Lofi beats to study and relax to.”
Most of the frame is taken up by a computer monitor, where the model grinds away at long-running tasks. The humans — who by rights would otherwise be pinned to a desk — are off fishing, playing tennis, scaling rock faces.
One company wants to sell you protection from its own creation. The other wants to sell you an afternoon off.
The polling behind the tonal split
Chinese public sentiment toward AI runs noticeably warmer than American sentiment, and that probably accounts for a good share of the difference in pitch.
A 2023 survey conducted by accounting firm KPMG International together with the University of Queensland took the temperature in seventeen countries. Respondents in China came back the most positive of any group: 95% described themselves as optimistic, while 36% said they were “fearful.”
In the U.S., just 41% of respondents said they thought AI’s benefits outweigh its risks.
Domestic surveys keep landing in the same place. Americans are, for the most part, lukewarm at best about AI seeping into workplaces, social media, politics and everything else — hardly the enthusiasm Silicon Valley had banked on.
What to watch next week
The open weights release is the event that counts. The moment Qwen3.8-Max is available, anyone can run the visual reasoning and agentic computer use tests firsthand instead of accepting Alibaba’s word on the two victories it’s claiming.
Should those results survive scrutiny at $2 and $6 per million tokens, no mood-piece commercial is going to be much of a rebuttal to the pricing argument.