Six months ago, the gap between open-source models and GPT-4 class models was a chasm. Today it’s a crack. And for a lot of use cases, it’s gone entirely.

Llama 3, Mixtral, and the latest wave of fine-tuned open models are hitting performance levels that would have been unthinkable a year ago. For founders building AI products, this changes the math on build vs. buy in fundamental ways.

What Changed

Three things converged: better training data (thanks to synthetic data pipelines), more efficient architectures (mixture of experts went mainstream), and a flood of compute from companies competing to host open models.

The result: for tasks like summarization, classification, extraction, and simple reasoning, open-source models are now competitive with or better than proprietary APIs — at a fraction of the cost.

What This Means for Builders

If you’re building an AI product today, the default should no longer be “call OpenAI.” The decision tree looks more like:

• Need cutting-edge reasoning? Proprietary models still lead here.
• Need consistent, predictable outputs? Fine-tuned open models often win.
• Need to run on-premise or in regulated environments? Open source is your only option.
• Optimizing for cost at scale? Open source can be 10-50x cheaper.

The New Playbook

Smart startups are adopting a hybrid approach: prototype with proprietary APIs, then migrate performance-critical paths to fine-tuned open models once the use case is validated. This gives you speed in the beginning and economics at scale.

The companies that build abstraction layers making this migration easy will have a significant advantage over those locked into a single provider.

What to Watch

Keep an eye on the inference optimization space. The bottleneck for open-source models is no longer quality — it’s serving cost and latency. Companies like Groq, Together, and Fireworks are competing aggressively here, and the winner will determine how accessible these models become.