Open Weights vs. Proprietary AI: The Gap Stays at 42 Points

The Mode read

The open-weight versus proprietary debate reached its historic inflection in early 2025, when DeepSeek-R1 briefly overtook every proprietary frontier model, scoring 9 points higher than the best closed API on Chatbot Arena. That moment disproved the assumption that proprietary labs had an unbridgeable scaling lead.

Since then, frontier labs have reclaimed a modest lead. Across 2026, proprietary flagships have maintained a steady gap of 42 points over open weights. On an Elo scale, a 42-point gap translates to a 56% to 44% win rate on blind human preference—statistically noticeable, but functionally indistinguishable across routine business workflows like summarization, extraction, and first-tier customer support.

For venture founders and CTOs, a 42-point gap is not a barrier; it is an economic green light. Running open-weight models self-hosted or via zero-margin inference providers cuts token costs by 80% to 95%. When baseline competence is ubiquitous and inexpensive, product margins expand for applications that own proprietary distribution channels.

A 42-point gap does not justify a 10× price premium for 80% of workflows.

What we would back: Sovereign enterprise infrastructure that runs open weights locally, bypassing centralized API rent.