As open-source AI models continue to gain momentum, Anthropic CEO Dario Amodei argues that the term “open source” does not carry the same meaning in artificial intelligence as it does in traditional software development.
According to Amodei, most AI models are not truly open source because developers cannot inspect or understand how the model actually works. Instead, what is often released are the model’s weights, which is why many researchers prefer the term “open weights” rather than “open source.”
He believes that this distinction matters because the collaborative advantages traditionally associated with open-source software do not fully apply to modern AI models. While thousands of developers can contribute to improving an open-source software project, large language models cannot be developed in the same additive way simply by making their weights publicly available.
For Amodei, the real question is not whether a model is open or closed, but whether it performs well. “When I see a new model come out, I don’t care whether it’s open source or not. I ask: Is it a good model? Is it better than us at the things that matter? That’s the only thing that I care about.”
He pointed to DeepSeek as an example, arguing that its success is driven by the quality of the model rather than by the fact that it is openly available. Amodei also challenged another common assumption: that open AI models are essentially free. Even when model weights are publicly accessible, running modern AI systems requires expensive infrastructure, powerful GPUs, and cloud-based inference services.
In other words, making a model available does not eliminate the costs of deploying and serving it at scale. Someone still has to provide the computing power that makes the model fast, reliable, and usable.
Amodei’s remarks suggest that competition in the AI industry is gradually shifting away from the debate over “open source” versus “closed source.” Instead, success will increasingly be determined by practical performance—how well, how fast, and how efficiently a model solves real-world tasks.
