The headline benchmarks still belong to the frontier labs. But most work inside a company is not a benchmark. It is classifying a message, drafting a reply, extracting a field, summarising a thread. For that long tail, open-weight models quietly became good enough this year, and "good enough plus private plus cheap" is a strong offer.
Where open now wins
- High-volume, well-defined tasks where the cost per call matters more than the last few points of quality.
- Anything with data that cannot leave your walls, where running the model yourself is the whole point.
- Products that need predictable pricing and no dependency on a single vendor's roadmap.
Most software does not need the smartest model in the world. It needs a reliable one it is allowed to use.
Where the frontier still matters
For the genuinely hard reasoning, the long-horizon agent work and the tasks where a mistake is expensive, the gap is still real. The emerging pattern is not "open versus closed" but a portfolio: open models for the volume, frontier models for the hard edge, and a router deciding which gets what.
Worth your time
- A practical guide to routing between open and frontier models.
- Designing Data-Intensive Applications, still the reference when you start hosting your own.