How far behind is open?
Some AI models you can only rent — you send text to a company's server and pay per use, and you never hold the model itself. GPT, Claude and Gemini work this way. Others you can download: the finished model is a file you run on your own hardware, privately, for the cost of the machine. Llama, DeepSeek and Qwen work this way.
The downloadable ones are behind. This page answers by how much — not in points on some index, but in time: the best model you can download today is about as good as the best rentable model was some months ago. That number is the single most-argued figure in the field, and nobody publishes it as a series, which is why we store every reading rather than recomputing it on the page.
6.3 months
counting weights you can download, at the best estimate.
The best model you can download and run yourself (Kimi K3) is about as good as the best rented model was 6.3 months ago — that is GPT-5.4 Pro, from March 2026. Count it differently — a stricter idea of "open", or the cautious end of the measurement error — and it lands anywhere between 4.7 and 9.1 months, so treat it as a season rather than a number.
Too little history to call a direction — the readings so far span under a month.
Every method we run4.7–9.1 months
| Open means | Read at | Gap | Best open model | Its 95% interval | Closed passed it |
|---|---|---|---|---|---|
| weights you can also ship commercially | if open is at the top of its interval | 6.3 months | DeepSeek V4 Pro 0813 · DeepSeek | 95% CI 153.8-157.8 | GPT-5.4 Pro · 2026-03-05 |
| weights you can also ship commercially | if open is at the bottom of its interval | 9.1 months | DeepSeek V4 Pro 0813 · DeepSeek | 95% CI 153.8-157.8 | GPT-5.2 Pro · 2025-12-11 |
| weights you can also ship commercially | best estimate | 7.3 months | DeepSeek V4 Pro 0813 · DeepSeek | 95% CI 153.8-157.8 | GPT-5.3 Codex · 2026-02-05 |
| weights you can download | if open is at the top of its interval | 4.7 months | Kimi K3 · Moonshot | 95% CI 155.4-160.4 | GPT-5.5 Pro · 2026-04-23 |
| weights you can download | if open is at the bottom of its interval | 9.1 months | Kimi K3 · Moonshot | 95% CI 155.4-160.4 | GPT-5.2 Pro · 2025-12-11 |
| weights you can download | best estimate | 6.3 months | Kimi K3 · Moonshot | 95% CI 155.4-160.4 | GPT-5.4 Pro · 2026-03-05 |
How it is measured
Both sides are running maxima of Epoch AI's Capabilities Index over release date — the best anyone had shipped by that day, never falling when a weaker model appears. The gap is the time between today and the day the closed frontier first reached the level the best open model is at now.
It is dated by release, not evaluation, and that is the one place on this site where release date is the right choice. Everywhere else we date by evaluation, because the gap between the two is where contamination hides. Here the question is when a capability became available to you, which is a fact about shipping.
Two choices move the answer, and neither is obvious, so we publish both rather than pick one quietly. “Open” can mean weights you can download or weights you can also ship commercially — different questions with different answers. And every index score carries a 95% interval roughly seven points wide, comparable to a year of frontier progress, so reading the gap off the midpoint alone would claim a precision the source does not.
Models whose access Epoch could not establish are excluded from both sides. Putting an unknown model on a frontier it may not belong on would move the number in a direction nobody could check.
Source: Epoch AI, “AI Benchmarking Hub”, used under CC-BY 4.0. Attribution is a licence condition, not a courtesy. Today's reading compares 126 open-weights models against 118 closed ones.