AI release tracker puts fresh dates on 20 models as Chinese usage surges

The page is called "How stale is your AI?" and it does something most launch posts do not: it separates the day a model shipped from the day its training data stopped.

The tracker sits at stale.jock.pl and was posted to Hacker News on 16 September. It covers 20 models from 8 labs, with live counters running from each date. Ten of the 20 carry a training cutoff the lab actually publishes. The gap between the two dates is the point. GPT-6 Astra, for example, was released on Sep 3, 2026, and its training data stops at Apr 30, 2026, so the model was four months behind on day one, the page says. Documentation quality varies across the list. Anthropic, Google DeepMind, Meta, OpenAI and xAI have a published cutoff for at least one model on the page, according to the tracker, which is 5 of 8 labs. Mistral AI and Alibaba models listed there show "Not established," which the page is careful to describe as a blank in the checked vendor sources, not proof that no cutoff exists.

"A model can ship in September and still stop reading in April, which means it is five months behind on the day it launches."

The page also reports an experiment on search behaviour. The author ran over 2000 calls across 16 models, each given a web search tool. The frontier models decided correctly almost every time. Weaker ones answered settled questions from memory after the answer had changed, and searched the web for things like the boiling point of water. The author's summary of the failure mode is blunt: "I gave 16 AI models a search button and asked who the king of Norway is. Five named a dead man."

Usage data points the other way

Freshness is one framing. Adoption is another, and the numbers there are moving faster. Chinese AI models went from a small share of usage to a majority on two major developer platforms that route traffic to different providers, according to usage data shared with CNBC and reported on 26 September. On OpenRouter, Chinese models accounted for 57% to 67% of tokens used in the week of Sept. 14, up from 6% to 13% in February. On Vercel, their share rose to 55% in August from 11% in January.

OpenRouter's figures cover companies in the U.S., Europe and what it defines as the "Global South," 82 countries across Central and South America, Africa and Asia. Vercel did not specify a geographical breakdown, CNBC said. Businesses in the Global South have been the biggest users of Chinese models on OpenRouter's system in recent weeks, with 67% of their tokens going to Chinese models, while about half the tokens on the platform come from U.S. companies.

Price is the stated driver. Peter Walker, head of insights at OpenRouter, told CNBC that Chinese open source models released this year "can credibly perform in advanced agentic use cases, especially in regards to coding, in a way that was just not true in late 2025," and that they are "incredibly cost-effective compared to most models from American labs." Harpreet Arora, head of agentic infrastructure at Vercel, said Chinese models are "becoming capable enough for more tasks at a much lower cost."

The political layer is thickening. CNBC reports that two U.S. House Committees are investigating the impact of rising adoption of Chinese models, and that AI was a major focus when President Donald Trump and President Xi Jinping met this week. Daniel Remler, a senior fellow at the Center for a New American Security, told CNBC the "ultimate concern is that the integration of Chinese AI models pulls countries into a Chinese technology sphere of influence that hardens into geopolitical alignment."

Separately, enterprise vendors are still shipping on top of U.S. open weights. Salesforce announced Koa on 15 September, a CRM reasoning model built by post-training NVIDIA Nemotron 3 Super on a synthetic dataset modelled on nearly three decades of CRM deployments. Salesforce says Koa matches or exceeds leading model performance on its CRM benchmark with three times fewer errors, and that it controls the weights and runs inference inside its own trust boundary. The model is moving into customer pilots with 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine and Xero.

Back on the tracker, the practical advice is aimed at agents rather than people: paste a few lines into the AGENTS.md or CLAUDE.md file an agent already reads, telling it to fetch models.json before naming any model, version or date as current. The export carries release date, published cutoff and a source link per model. "A model is a poor source on models," the page says.