Eighteen months ago the conventional wisdom was that open-source models would always trail flagship closed models by a year or more. The gap, the thinking went, was compute — and compute was only getting more expensive.

The conventional wisdom was wrong.

What actually happened

Better training recipes compressed the gap faster than raw compute widened it. A 70B open model trained on a well-curated dataset now matches last year's frontier models on most practical benchmarks. Fine-tuned variants beat them on specific tasks.

For builders, the practical implication is a real choice again. Host your own for cost and privacy, or call an API for the last few percent of quality. Both are viable. A year ago only one was.