Two races got faster this week, and they are not the same race. One is about who can give the most model away for free. The other is about who reaches the public markets first.
Alibaba released the weights of Qwen3.8-Max, a 2.4 trillion parameter model with 95 billion active, and shipped a 27B version small enough for a laptop. That pushes the open-weight lead further toward Chinese labs. Simon Willison tried the small one and found it genuinely good but strangely eager to think, burning minutes on questions that need seconds, so turn the reasoning setting down before you judge it. Nathan Lambert reads Nvidia’s 26 billion dollar push into open models as a bet that everyone keeps building instead of buying from Anthropic or OpenAI, and he is not convinced open weights can ever pay that bill.
The buying side moved too. Stripe agreed to acquire OpenRouter for around 7.5 billion dollars, which turns model routing from a developer convenience into payment infrastructure. Anthropic’s annualized revenue passed 65 billion dollars, more than seven times where it sat at the end of last year, while OpenAI’s quarterly growth slowed to 18 percent just before its own planned listing. Two companies, same market, very different graphs.
Research had a harder message for anyone shipping agents. A paper called CompoSkill shows that agent skills which each pass a safety scanner on their own can be chained into harmful combinations, with risk chains forming in over 80 percent of attempts, because the danger lives in the path and not in any single step. A second group pulled hidden reasoning out of black box models through a replay surface between tool calls, recovering traces almost word for word. Both point the same way. Safety work is moving off the model and into the plumbing around it.
One better result to end on. Training a generator against a critic judged by a deliberately weaker judge closed a 45 percent performance gap that reward hacking had opened, and it held across many training steps. Does that survive contact with a real agent stack?
T.
