Two years ago, the conventional wisdom was that open-source AI would always lag behind proprietary models. The reasoning was straightforward: only companies spending billions on training compute could produce frontier models, and they’d never give those models away.
That argument is dead. Meta’s Llama 4 405B matches GPT-4o. Mistral’s models regularly top open leaderboard rankings. Together AI, Fireworks, and other inference providers have made running open models cheaper than calling proprietary APIs.
The Economic Argument
The cost of training a frontier model has dropped from an estimated $100M+ in 2022 to under $10M in 2025, thanks to better training techniques (mixture of experts, distilled training, synthetic data), cheaper hardware, and more efficient architectures.
At $10M per training run, the economics change dramatically. A well-funded startup can afford multiple training runs per year. Academic labs can compete. The “compute moat” that was supposed to protect proprietary AI companies is eroding.
Why Meta Is Winning This Game
Meta’s strategy is particularly clever. By releasing open models, they commoditize the AI infrastructure layer that their competitors depend on. If any company can run a GPT-4o-class model for $0.50/M tokens instead of $15/M tokens, the proprietary API business becomes much less attractive.
Meanwhile, Meta benefits from the ecosystem effects: more developers building on Llama means more applications, more feedback, more fine-tuned variants, and a larger community that indirectly improves Meta’s own products.
The Caveat
Open source doesn’t win everywhere. The very largest models (1T+ parameters) are still within reach of only the biggest labs. And frontier capabilities — the bleeding edge of what’s possible — will likely remain proprietary for the foreseeable future. But the gap is narrowing, and for 95% of use cases, open source is already good enough.