Open-weight models, whose weights anyone can download, run and adapt, have long trailed the best closed models from the big labs. How big is the gap today? We compared the 61 open and 75 closed models that have an independent score. For the basics, see our guide to open and closed models.
The gap at the top
The best open model, Kimi K3 from Moonshot AI, scores 157.7 on the overall index. The best closed model, GPT-6 Astra from OpenAI, scores 166.6: a gap of 8.9 points. The best open model ranks #12 of the 136 rated models, and it is the only open model in the top 20.
The price gap is much bigger
Where open models shine is price. The median open model costs $0.40 per million tokens, against $3 for the median closed model (blended prices: three tokens read for each token written). Kimi K3 itself costs 3.3× less than GPT-6 Astra. Open models are served by many competing hosting companies, which keeps their prices low.
| Open model | Company | Overall index | Blended price |
|---|---|---|---|
| Kimi K3 | Moonshot AI | 157.7 | $6 |
| GLM 5.3 | Z.ai (Zhipu) | 155.6 | $2.15 |
| DeepSeek V4 Pro 0813 | DeepSeek | 155.4 | $0.99 |
| DeepSeek V4.1 Flash | DeepSeek | 155.0 | $0.52 |
| DeepSeek V4 Flash 0731 | DeepSeek | 154.5 | $0.094 |
| GLM 5.3 Flash | Z.ai (Zhipu) | 151.9 | $0.24 |
What it means for your choice
- If you need the very best results today, closed models still lead, by 8.9 points at the top.
- For most production work, the leading open models come close for a fraction of the price: Kimi K3 (157.7), GLM 5.3 (155.6), and DeepSeek V4 Pro 0813 (155.4).
- Open weights give you options that closed models cannot: keep your data on your own servers, fine-tune the model on your data, or change hosting provider without changing the model.
- Hosting a model yourself has its own costs (hardware, engineering, monitoring). The prices above are what API providers charge.
The best open models page ranks them all, with their prices and providers.
The figures in this article are recalculated from our data at every update (last: Sep 30, 2026). How we work