# How far behind is open-source AI?

> Source: <https://yaroslavvb.github.io/artificial-analysis-oss-lag/>
> Published: 2026-06-18 15:30:53+00:00

The open-source Pareto frontier over time: each point is an open-weights model that beat every open model before it — plotted by how many months earlier a proprietary model reached the same level.

The vertical position is the gap to the *earliest proprietary model with the same Artificial Analysis
Intelligence Index* (hover any point to see which one). The gap peaked near 10 months around DeepSeek V3
(Dec 2024) and has since tightened to ~2–3.5 months as DeepSeek, Kimi, Z AI and MiniMax began
shipping close behind the frontier.

**Frontier-pusher.** A model is included only if its Intelligence Index exceeded every open-weights
model released before it. 23 qualify. (Solar Mini, which the naive running-max flags in early 2024, is excluded: at
floor-level scores the retroactive v4.1 index ranks it above contemporaneous stronger open models — Mixtral 8×7B,
Llama-2-70B, Qwen-72B — so it was not actually the leading open model at release.)

**Metric.** Artificial Analysis Intelligence Index (v4.1), applied consistently to every model across
all dates. It is anchored to today's hard benchmarks (GPQA, HLE, terminal-bench, etc.), so 2023 models score low and
their ordering at the floor is noisy.

**Gap.** For a frontier model released on date `D`

with index `S`

, gap =
`D − T`

, where `T`

is the earliest date any proprietary model reached index
`≥ S`

— that proprietary model is shown in each point's tooltip.

**Note.** This is the gap at a given *capability level*, not the absolute frontier. GLM-5.2
(II 51) matches GPT-5.4 from 3.4 months earlier, while the absolute closed frontier (Claude Fable 5, II 60)
sits further ahead in raw capability.