Open AI models trail the frontier by only months — is that just because GPUs keep getting better, or is something else closing the gap?

This explores whether open models stay roughly 6–12 months behind the frontier mainly because consumer hardware keeps getting more powerful, or whether other forces do most of the catching up. The short answer is no. Hardware is one of three forces the corpus points to, and the other two may matter more. Epoch AI found that open models small enough to fit on a single high-end consumer GPU match frontier benchmark scores from 6–12 months earlier. They attribute that lag to three things together: open and frontier labs scaling at similar rates, distillation (training small models to imitate big ones), and better GPUs How quickly do open models catch up to frontier AI?. The gap also isn't fixed. On LM Arena, these open models gain Elo about 56% faster than frontier models, so the gap is shrinking Are open models closing on frontier AI faster than expected?. If hardware were the main driver, you'd expect the two groups to move in step, not one closing in on the other.

The strongest counter-evidence to a hardware-only story comes from training methods. A 3B-parameter model, trained with carefully ordered fine-tuning examples (curriculum SFT) and then reinforcement learning across several domains, matches much larger systems on competition math and coding benchmarks Can small models match frontier reasoning without massive scale?. The catch is that this only holds where answers can be checked automatically. In those areas, though, recipe beats raw size. A related finding is 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?. That suggests 'frontier capability' isn't fixed in the weights at all.

A third force has nothing to do with the model itself: the harness, meaning the software system wrapped around a model. Improving how the model runs tools and keeps state lifts accuracy on Terminal-Bench without changing any weights, and the same setup transfers to open models like DeepSeek Can execution harnesses lift model performance without retuning weights?. Layered memory systems, from weights out to disk-backed history, push this further Can external state caches let models solve harder problems?. Some researchers argue that how memory is organized now pays off more than adding parameters Has memory architecture replaced parameter count as the scaling frontier?. The most striking example: with better scaffolding, Llama3.1-70B and Qwen2.5-72B copied themselves in tests where GPT-o1 failed. Success tracked how much the model explored, not raw capability Can weaker open-source models achieve self-replication?.

There's also a twist in what 'frontier' even measures. Roughly half of recent benchmark gains on GPQA-Diamond track higher inference spending, not gains in the model itself. Price-performance improves 5–10× a year, but the absolute cost of running frontier models rises 3–18× a year Is benchmark progress outpacing the cost to achieve it?. So part of the frontier's lead is bought at run time, and an open model with good scaffolding and more thinking time can buy some of it back.

The unexpected takeaway is that the gap may be less about who owns the chips and more about who has the best training recipes and harness designs, and both of those spread quickly. One limit: the corpus doesn't have a study that separates how much each of the three forces contributes, so their relative weights are still an open question.