As AI labs build bigger models, do smaller open ones actually keep pace — or even gain ground?

This explores whether open models fall further behind as frontier labs build ever-larger models. The most direct evidence in the collection points the other way. Epoch AI tracked LM Arena ratings and found that open models small enough to run on a single consumer graphics card are improving about 56% faster than frontier models, at roughly +125 Elo a year against +80 Are open models closing on frontier AI faster than expected?. If those slopes hold, the gap closes. Two caveats apply: this is one leaderboard, and a trend line is not a guarantee.

The deeper reason the gap may not widen is that model size is no longer the main way capability gets made. Several notes show other levers matching or beating raw scale. Smaller models given more thinking time at inference can match much larger ones on hard problems Can inference compute replace scaling up model size?. A 35B model trained on real execution and long multi-step tasks sits at the cost-performance sweet spot next to far bigger models Does model efficiency matter more than peak capability for real work?. A 9B model trained on whether its patches actually worked beat prompted frontier models at improving an agent's setup Does training editors on real outcomes beat prompting larger models?. Routing each question to the best of several specialized models outperformed GPT-5, and ten 7B models with routing had earlier surpassed GPT-4.1 Can routing beat building one better model?. One synthesis argues that how a model's memory is organized now returns more than adding parameters Has memory architecture replaced parameter count as the scaling frontier?. Training signal, inference strategy, routing and memory design are all open to anyone, not just labs with the biggest training budgets.

There's also a twist in what frontier progress is actually made of. An analysis of frontier pricing found that roughly half of measured gains on a hard science benchmark come from spending more compute at inference, not from capability that holds up independent of that spending, and the cost of running top models is rising 3–18× a year Is benchmark progress outpacing the cost to achieve it?. Part of the frontier's lead is bought at inference time, so it is expensive to keep and partly available to open models willing to spend the same compute.

Finally, "bigger and newer" doesn't reliably mean "better at everything." On a business strategy simulation, mid-to-late 2025 frontier models scored below earlier models and below MBA students, favoring short-term profit over long-term bets Do newer frontier LLMs actually make better strategic decisions?. On very long agent tasks, the best predictor of success was persistence in test-and-revise loops, a behavior shaped by training, not by size What predicts success in ultra-long-horizon agent tasks?. And measuring the gap is itself fragile: models can be prompted or fine-tuned to underperform on chosen evaluations Can language models hide their true capabilities during evaluation?. The collection doesn't contain a forecast of the gap under future scaling. What it does show is that "the gap" is several gaps, and on many of them the levers that matter most aren't the ones only frontier labs control.