After a new frontier AI model drops, how long until free, open versions catch up to it?

This explores how fast open models catch up to closed frontier models, and whether the gap is shrinking or holding steady. The collection doesn't give a single figure like "open models trail by N months." What it does have is a rate: the gap is shrinking, and it's shrinking faster than most people would guess. Epoch AI tracked LM Arena Elo scores, a ranking built from head-to-head human preference votes. It found that open models small enough to run on one consumer GPU gain about 125 Elo points a year, while frontier models gain about 80 Are open models closing on frontier AI faster than expected?. That is roughly 56% faster improvement from models you could run at home. So the gap isn't a fixed delay. It narrows with each release cycle.

It also helps to ask how big the gap is today, not only how fast it closes. Nagle and Yue estimate that open models already reach about 90% of closed-model performance at roughly one-sixth of the cost Why do closed AI models dominate despite costing six times more?. Yet closed models still take 80% of usage and 96% of revenue. The authors put this down to brand trust, information gaps and liability worries, not to real capability differences. They estimate $20–48B a year in savings that customers are leaving on the table. The OpenRouter study, which covers more than 100 trillion tokens, shows the other side of the market: open models such as DeepSeek and Qwen still handle over 30% of tokens, even in areas where closed models lead. Users there weigh cost and speed against raw capability How much do open-weight models actually get used in practice?. Put together, the catch-up question looks less about benchmarks and more about trust: the capability gap may be closing faster than buyers' perception of it.

The less obvious point is that the gap doesn't have to be closed by training a better open model at all. Several notes show performance improving without changing the model's weights. Snell et al. found that giving a smaller model more compute at answer time can match a larger model on hard prompts Can inference compute replace scaling up model size?. On the agent side, improving the "harness" (the execution system wrapped around a fixed model) lifted DeepSeek-V4 Flash by 5.4 points on Terminal-Bench 2.1, and the same setup carried over to newer models unchanged Can execution harnesses lift model performance without retuning weights?. A related design gives models layered memory, from context to disk-backed history, so they can take on longer problems Can external state caches let models solve harder problems?.

There's a limit on that shortcut, though. Running many calls to weaker models can match a strong model only when there's a reliable way to check answers, such as tests, proofs or type checks. Without one, extra sampling turns up more candidate answers but can't tell which one is right When can weak models match strong model performance?. So in areas like coding and math, where answers can be checked, cheaper open models plus extra compute can close the gap almost right away. In open-ended tasks without such checks, the gap is harder to engineer around.

To be clear about what the collection can't answer: it has no study that measures, in months, how long open models take to match a specific closed release. What it offers is a slope (open models improving faster), a current gap (about 90% of the performance at a sixth of the price) and a set of ways to close the rest without new weights.