Every year the "open models are two years behind" claim gets re-tested, and every year the gap shrinks. In August 2026, the best open-weights model we track — Kimi K3 from Moonshot AI — holds a Podium Score of 83.3, while the best proprietary model, Claude Mythos Preview, sits at 97.4.
The gap in one number
On our normalized 0–100 scale, the open-weights frontier is about 14 points behind the proprietary frontier. That is roughly the distance between the #1 and #8 proprietary model — significant, but no longer a different league. On LiveCodeBench, Kimi K3 (74.7%) beats several closed models that cost 10× more per token.
Top open-weights models right now
- Kimi K3 (Moonshot AI) — 83.3
- GLM-5.2 (Z.ai) — 59.1
- MiMo V2.5 Pro (Xiaomi) — 50.3
- DeepSeek V4 Pro — 44.8
- Kimi K2.6 — 44.3
The long tail drops off quickly: after the top three, open-weights scores fall below 50. The honest reading is that one open lab currently competes at the frontier, while the rest cluster a tier below.
When open weights win anyway
Self-hosting, data privacy, fine-tuning and per-token economics. DeepSeek V4 Flash at $0.28/M output tokens makes high-volume workloads viable where $25–50/M proprietary pricing does not. For regulated industries the ability to run weights on-prem is often worth more than a few benchmark points.
Bottom line
The 2026 gap is real but narrow at the very top. Track it live on the Open-Weights Leaderboard and compare any pair on the comparison tool.