For most of the generative-AI era, the answer to “should I use an open or a closed model?” was simple to the point of being boring. If you wanted the best, you paid OpenAI, Anthropic, or Google, and open models were the respectable budget tier — fine for tinkering, not for anything that mattered. That answer is now out of date, and the data explaining why is worth looking at directly rather than through the fog of vendor marketing on both sides.
Start with the single most-cited figure. According to the Stanford HAI AI Index 2025, the performance gap between the best closed model and the best open-weight model on the Chatbot Arena leaderboard shrank from eight percentage points in January 2024 to just 1.7 points a year later. That is not a gradual narrowing. That is a gap collapsing inside twelve months. On broad knowledge tests the convergence is now so complete that a benchmark like MMLU barely separates anyone anymore, because nearly every capable model, open or closed, scores above 90.
If you stopped reading there, you would conclude the debate is over and open source has won. But stopping there would be exactly the kind of half-story that makes this topic so confusing, because a second, equally credible source tells you the gap is real and still measured in months.
Two true things that sound contradictory
The research group Epoch AI tracks this differently — not as a percentage on one leaderboard, but as a time lag: how long does it take the best open-weight model to catch up to where the closed frontier already is? By Epoch’s measurement, open models trailed the frontier by about three months on average between early 2023 and late 2025, and that lag has widened slightly to roughly four months since the start of 2026.
So which is it — a gap of 1.7 percent, or a gap of four months? The answer is both, and holding the two together is the whole point. A benchmark score measures distance at a single moment; a time lag measures how fast the frontier is running away. Open models can be within a whisker of today’s best scores while still being a season behind the newest closed release, because the closed frontier keeps moving. Parity on last quarter’s tests is not the same as parity with what just shipped.
Epoch adds two caveats that are easy to miss and genuinely important. First, open-weight models may look better than they are, because they tend to be tuned aggressively on public benchmarks — the very tests everyone quotes. On private benchmarks they tend to do worse. Second, the closed labs do not always release their most capable systems, for safety, commercial, or competitive reasons. Both factors mean the true gap is probably a little wider than the tidy public numbers suggest. That is the opposite of a marketing message, which is exactly why it is worth trusting.
Where each side actually wins
Once you accept that the honest answer is “it depends,” the useful question becomes: depends on what? Here the picture is clearer than the headline war implies.
Closed frontier models still hold a measured lead where it is hardest to fake: the hardest reasoning problems and long-horizon agentic tasks. The picture on everyday coding has actually shifted — on the widely watched SWE-bench Verified leaderboard the top open-weight models now sit within a point or two of the best closed ones, and on some coding evaluations an open model leads outright. But the lead reasserts itself on the messier, multi-step problems open models train hardest against; on one long-context coding benchmark, open models managed only single-digit solve rates while the best closed model still cleared barely a fifth. If your use case is high-stakes and capability-bound — the work where the difference between the 95th and 99th percentile answer actually costs money — the closed frontier is still the safer bet.
Open-weight models win on a different axis entirely, and it is not just price, though the price gap is enormous. They can be fine-tuned on your own data, run on your own hardware, deployed in an air-gapped environment where nothing leaves the building, and audited in ways a closed API simply does not permit. For a hospital, a bank, or a school district weighing whether “the good models are expensive and send our data to someone else’s servers” is still a reason to hold back, that trade-off has genuinely changed. The near-frontier is now available under a license you control.
The pattern most teams are actually settling on
The real-world resolution to the debate turns out not to be a winner at all. It is a division of labor. A common production setup routes the large majority of requests to a cheaper open model and escalates only the hardest fraction to a frontier closed model — capturing most of the cost savings while keeping the capability ceiling available for the requests that need it. The question stopped being “open or closed” and became “which request goes where.”
There is a bigger structural shift underneath all of this that the Stanford data captures almost in passing: no single model dominates anymore. The report found the spread between the best and tenth-best model narrowing from 11.9 percent to 5.4 percent in a year, with the top two sometimes separated by less than a point. A landscape with four or five genuinely competitive labs, plus a fast-following open ecosystem, is one where pricing power erodes and buyers gain leverage. Whatever you think about the open-versus-closed scoreboard, that competition is the development most likely to matter for anyone actually paying the bills.
The tidy narratives — “closed models are unassailable,” “open source will replace everything by summer” — are both wrong, and comfortably so. The accurate version is less quotable and more useful: the capability gap has shrunk to a crack, the deployment trade-offs have not, and the smart move in 2026 is to stop rooting for a side and start matching the model to the job.
Related reading: Everyone’s Adopting AI. Almost Nobody’s Profiting. That’s the Real 2026 Story
