In today’s AI race, breakthroughs are no longer measured in years—or even months—but in weeks.
The release of Opus 4.6 just over two weeks ago was a major moment for its maker, Anthropic, delivering state-of-the-art performance in a number of fields. But within a week, Chinese competitor Z.ai had released its own Opus-like model, GLM-5. (There’s no suggestion that GLM-5 uses or borrows from Opus in any way.) Many on social media called it a cut-price Opus alternative.
But Z.ai’s lead didn’t last long, either. Just as Anthropic had been undercut by GLM-5’s release, GLM-5 was quickly downloaded, compressed, and re-released in a version that could run locally without internet access.
Allegations have flown about the ways AI companies can match, then surpass, the performance of their competitors—particularly how Chinese AI firms can release models rivaling American ones within days or weeks. Google has long complained about the risks of distillation, where companies pepper models with prompts designed to extract internal reasoning patterns and logic by generating massive response datasets, which are then used to train cheaper clone models. One actor allegedly prompted Google’s Gemini AI model more than 100,000 times to try and unlock the secrets of what makes the model work so powerfully.
“I do think the moat is shrinking,” says Shayne Longpre, a PhD candidate at the Massachusetts Institute of Technology whose research focuses on AI policy.
The shift is happening both in the speed of releases and the nature of the improvements. Longpre argues that the frontier gap between the best closed models and open-weight alternatives is decreasing drastically. “The gap between that and fully open-source or open-weight models is about three to six months,” he explains, pointing to research from the nonprofit research organization Epoch AI tracking model development.
The reason for that dwindling gap is that much of the progress now arrives after a model ships. Longpre describes companies “doing different reinforcement learning or fine tuning of those systems, or giving them more test time reasoning, or enabling to have longer context windows”—all of which make the adaptation period much shorter, “rather than having to pre-train a new model from scratch,” he says.
Each of those iterative improvements compounds speed advantages. “They’re pushing things out every one or two weeks with all these variants,” he says. “It’s like patches to regular software.”
But American AI companies, which tend to pioneer many of these advances, have become increasingly outspoken against the practice. OpenAI has alleged that DeepSeek trained competitive systems by distilling outputs from American models, in a memo to U.S. lawmakers.
Even when nobody is “stealing” in the strict sense, the open-weight ecosystem is getting faster at replicating techniques that prove effective in frontier models.
The definition of what “open” means in model licenses is partly to blame, says Thibault Schrepel, an associate professor of law at Vrije Universiteit Amsterdam who studies competition in foundation models. “Very often we hear that a system is or is not open source,” he says. “I think it’s very limited as a way to understand what is or what is not open source.”
It’s important to examine the actual terms of those licenses, Schrepel adds. “If you look carefully at the licenses of all the models, they actually very much limit what you can do with what they call open-source,” he says. Meta’s Llama 3 license, for instance, includes a trigger for very large services but not smaller ones. “If you deploy it to more than 700 million users, then you have to ask for a license,” Schrepel says. That two-tier system can create gray areas where questionable practices can emerge.
To compensate, the market is likely to diverge, MIT’s Longpre says. On one side will be cheap, increasingly capable self-hosted models for everyday tasks; on the other, premium frontier systems for harder, high-stakes work. “I think the floor is rising,” he adds, predicting “more very affordable, self-hosted, self-hosted, general models of increasingly smaller sizes too.” But he believes users will still “navigate to using OpenAI, Google and Anthropic models” for important, skilled work.
Preventing distillation entirely may be impossible, Longpre adds. He believes it’s inevitable that whenever a new model is released, competitors will try to extract and replicate its best elements. “I think it’s an unavoidable problem at the end of the day,” he says.