Open weights vs proprietary
Frontier AI capability, October 2025 to October 2026. Each month shows the best model available in each category, scored on the Intelligence Index (v4.3.2, applied retrospectively). Higher is better.
| Month | Proprietary | Open weights |
|---|---|---|
| Oct 25 | 23 | 18.5 |
| Nov 25 | 28 | 22 |
| Dec 25 | 30.4 | 22.2 |
| Jan 26 | 30.4 | 22.2 |
| Feb 26 | 30.4 | 22.9 |
| Mar 26 | 39 | 22.9 |
| Apr 26 | 40.7 | 24.2 |
| May 26 | 41.8 | 24.2 |
| Jun 26 | 49.6 | 33.7 |
| Jul 26 | 50.8 | 43.6 |
| Aug 26 | 50.8 | 44.8 |
| Sep 26 | 57.6 | 46.3 |
| Oct 26 | 57.6 | 46.3 |
Both frontiers rose about 150% over the year: open weights from 18.5 to 46.3, proprietary from 23.0 to 57.6. The current gap is 11.3 index points.
The particularly interesting finding is that open-weight models are improving at roughly the same proportional rate as proprietary models, although the absolute difference has widened over the year.
The gap is not constant, however. In November 2025, Kimi K2 Thinking narrowed it to approximately one index point. It widened dramatically in the following months, then narrowed again with Kimi K3 and GLM-5.3 during July and August 2026.
Which models moved the frontier?
| Date | Leading release | Type | Index score |
|---|---|---|---|
| Nov 2025 | Kimi K2 Thinking | Open-weight | 22.0 |
| Dec 2025 | GPT-5.2 | Proprietary | 30.4 |
| Mar 2026 | GPT-5.4 | Proprietary | 39.0 |
| Jun 2026 | GLM-5.2 | Open-weight | 33.7 |
| Jun 2026 | Claude Fable 5 | Proprietary | 49.6 |
| Jul 2026 | Kimi K3 | Open-weight | 43.6 |
| Aug 2026 | GLM-5.3 | Open-weight | 44.8 |
| Sep 2026 | Claude Opus 5.5 | Proprietary | 57.6 |
| Sep 2026 | MiMo-V2.6-Pro | Open-weight | 46.3 |
What this tells us
- Open-weight AI is not consistently closing the gap. It catches up in bursts, then a new proprietary release often reopens the distance. The overall trajectory is upward for both categories.
- The capability delay is more revealing than the raw score gap. The current leading open-weight score of 46.3 was first reached by a proprietary model approximately 3.4 months earlier.
- Open weights do not necessarily mean locally practical. The best models in this chart can be enormous mixture-of-experts systems requiring substantial memory and compute. A 27B or 70B model running on your own hardware may be considerably less capable than the open-weight frontier.
Source
Frontier: Open weights against the proprietary frontier, using Artificial Analysis data refreshed 24 October 2026. Scores are retrospectively evaluated under the same index version; some historical values are estimates. Monthly points show the best available model, not monthly averages.