Open AI Models Are Catching Closed Frontier Models in Half the Time, SemiAnalysis Finds

Open-weight AI models are closing the gap with closed frontier models at a striking pace: with each successive era of large language models, open models take half as long to catch up to the first closed model of that era, according to a new analysis by semiconductor research firm SemiAnalysis.

The study, titled "Are Open Models Catching Up?" and published August 21 by Evan Cloutier, Max Kan, Jordan Nanos, and Dylan Patel, is one of the most systematic attempts yet to quantify a question hanging over the entire AI industry: if open models stay nearly as capable as the closed frontier at a fraction of the cost, do frontier labs have a durable business? For deep coverage of the model releases driving this shift, follow AI Buzz Wire.

Three Eras of Frontier AI

The authors argue that the history of LLMs falls into three distinct eras — early scaling, reasoning, and agentic — each representing a step-function increase in what models can usefully do.

Their core methodological point is that comparing models across eras with a single benchmark is a mistake. Every benchmark is a product of its era: it is created to discern capability differences that exist at the time, then climbed by model makers until it saturates, at which point the industry stops caring and moves on. Instead, the SemiAnalysis team ran curated benchmark sets for each era's models to produce a composite capability score.

Viewed this way, the open-versus-closed gap moves in cycles. At the start of each era, a frontier lab completes promising research, trains an impressive model, and jumps ahead. Other labs then identify the key advances, reverse-engineer them, and close the gap. "Nothing stays secret forever — especially when you factor in distillation," the authors write. "It's just a question of how long it takes."

Their measurement of that question produced the study's headline finding: with each generation, open-source models take half as long to catch up to the first closed-source model of the era.

Why This Cycle Feels Different

The report is blunt about the difference between past open-model moments and the current one. The "DeepSeek moment" of January 2025 made headlines, but "no one actually used R1 to do any economically valuable work," the authors note.

By contrast, they point to models like GLM 5.3 and Kimi K3 as "genuinely capable of many of the same coding and agentic tasks that rocketed Anthropic to $65B+ ARR" — a reference to the revenue trajectory that coding and agentic workloads have driven for the Claude maker. That claim carries particular weight because these open models are not just demo-ware; they are being deployed in production for the same tasks that generate real revenue for closed labs.

The usage numbers reinforce the shift. Fireworks, an inference provider that serves many open models, is now processing more than 40 trillion tokens per day — twice the volume of OpenAI's API at the end of March, according to SemiAnalysis. It is, as the report puts it, "an exciting time to be a token consumer," with the battle for tokens now extending well beyond the OpenAI–Anthropic duopoly.

The Commoditization Question

The study directly addresses the fear — which it labels FUD from some corners — that capable, cheap open models will commoditize the model layer entirely, an outcome that would be disastrous for frontier lab margins.

The public portion of the analysis stops short of a full answer, pointing readers to the firm's paid Tokenomics Model for detailed treatment of OpenAI's and Anthropic's financials. But the direction of travel is clear: if each era's open models arrive in half the time of the previous era's, the window in which closed labs can charge premium prices for frontier capability keeps shrinking.

The authors also caution against reading the trend as a death sentence for frontier research. Benchmarks, they write, "don't tell the full story," and the report's conclusions are described as "less bearish" on frontier models than readers might initially think — in part because raw benchmark parity is not the same as parity in reliability, product integration, enterprise distribution, or the capital needed to serve models at scale.

Why It Matters

For buyers, the findings validate a strategy that many enterprises have already adopted: defaulting to open or cheaper models for the bulk of token volume while reserving premium closed models for the hardest tasks. For the labs, it quantifies the squeeze they face on two fronts — open models below, and enormous compute costs above.

And for the open-source ecosystem, the SemiAnalysis data offers something that has often been in short supply: evidence, rather than ideology, that the gap is closing on a predictable schedule. If the half-time pattern holds into the next era of frontier AI, the industry's most valuable question becomes not whether open models catch up, but what remains defensible after they do.

Track the Open-Weight AI Race Read more AI news →