The Open-vs-Closed AI Race Keeps Repeating Itself

Every major AI era begins the same way. A leading closed lab jumps ahead on a new capability, the field treats the gap as a moat, and within months a cohort of open-weight models catches up. Three cycles in, the question is no longer whether the gap will close. It is whether the closed-frontier business model survives the next closing.

That is the frame SemiAnalysis builds in its latest note, and it changes the story from "is open source winning?" into something more predictive: which era are we in, how catchable is the current lead, and what does the next compression do to the labs that priced the previous gap?

A quick doorway. "Frontier" labs are the handful of well-funded AI companies, including OpenAI, Anthropic, and Google DeepMind, that build the most capable models behind APIs or paid products. "Open-weight" or "open" models release their trained parameters publicly, so anyone can run or fine-tune them, usually far cheaper than the closed alternatives. "Closed" is everything else: the model is the product. With that vocabulary, the cyclical pattern becomes visible.

In the early scaling era of 2023, closed labs burned compute to win raw capability benchmarks. The gap looked structural. Within roughly a year, Meta's Llama family and a wave of open reproductions matched most practical performance at a fraction of the inference cost. The moat turned out to be a runway, not a wall. The reasoning era that followed, kicked off by OpenAI's o1-style chain-of-thought models in late 2024, looked like a reset. Then DeepSeek's R1 in January 2025 reproduced the reasoning pattern with open weights, and the cycle repeated. SemiAnalysis calls R1 a hype moment that did not translate into the economically valuable workloads many expected: useful, but not the threat the headlines suggested.

Today's era is the agentic one, where coding agents, multi-step tool use, and autonomous task execution are the workloads driving closed-frontier revenue. Anthropic now sits at over $65 billion in annualized run-rate revenue on SemiAnalysis's estimate, a figure the note treats as closer to reality than the inflated ARR claims circulating in the press, though still an analyst estimate rather than an audited disclosure. The bet is that agentic capability is durable enough to anchor pricing for years. That bet is now being tested by a new open-weight cohort.

Two open models, GLM 5.3 from Zhipu and Kimi K3 from Moonshot, are described in the note as genuinely capable of the same coding and agentic tasks that drove closed-frontier adoption. That is a sharper claim than "open models are improving." It is that open models can now do the work customers are paying $200 a month for. The third pole of the inference economy matters here too. Fireworks, an inference provider running open-weight models at scale, is processing more than 40 trillion tokens per day, roughly twice OpenAI's reported API volume at the end of March (SemiAnalysis estimate; OpenAI does not publish token throughput). A third of the inference economy is now flowing through open models, not the closed APIs.

If the cycle holds, the next catch-up is the test. Either the open-weight cohort closes the agentic gap the way it closed the reasoning gap, or the workload stays durable enough to protect closed-frontier pricing. SemiAnalysis is firm: this is a structural margin question, not a marketing one.

A useful way to read the three eras is not "open wins" or "closed wins." It is that the model layer keeps commoditizing faster than the labs can find a new capability to charge a premium for. Early scaling commoditized raw capability. Reasoning commoditized chain-of-thought. If agentic commoditizes next, the labs priced on durable agentic margins face the same compression that hit their predecessors. The SemiAnalysis piece names this directly: the central risk for frontier-lab economics is that open models stay capable at a fraction of closed-model cost. ARR figures like Anthropic's $65 billion-plus are not defensible if any developer can self-host a near-parity model and run it at inference prices a tenth of the API.

That is the live story. Not a "open source is winning" victory lap and not a "closed labs will pull ahead again" hedge. A falsifiable prediction: the agentic era's gap is the one to watch, and the watch is now.