Open-weight models are quietly closing the gap

Three open-weight releases this quarter landed within a few points of frontier closed models on standard benchmarks — a gap that was double digits a year ago.

That matters less for raw capability bragging rights and more for pricing pressure: self-hosting a model that's 95% as good, for a fraction of the API cost, changes the calculus for a lot of production workloads.

What actually closed the gap

Better training data curation and longer training runs on existing architectures did more work than any single new technique.

The frontier isn't standing still, but the distance behind it is shrinking faster than the distance ahead of it is growing.

What this means for builders

• Self-hosting is a realistic option for more workloads than it was last year
• Closed-model providers will feel pricing pressure on mid-tier models first
• The gap that remains is mostly in tool use and long-context reliability, not raw reasoning

Worth watching which way the frontier labs respond — price cuts, or a widening gap on the next generation.