This post throws out several very specific numbers: Anthropic’s rumored Opus is said to have reached 5 trillion parameters, OpenAI’s internal Mythos/Fable may be as large as 10 trillion, while open-source models stayed below 1 trillion for a long time. Our view is that this is not “breaking news,” but it does hit the most hard-headed layer of large-model competition: a lot of the lead may not come from mysterious tricks at all, but simply from who managed to push scale to the limit first.
What this is
The source is a Reddit discussion thread, not a company announcement and not a verifiable research report. The author’s core point is simple: OpenAI and Anthropic may not have the kind of “secret formula” outsiders imagine. Their real moat may be scale itself, including parameter count, training data, compute spend, and engineering execution. The post also mentions that after breaking through earlier parameter ceilings, models such as DeepSeek V4 and Kimi K3 showed a clear jump in capability.
Industry view
This line of thinking is not new inside the industry. Supporters argue that much of the progress in large models over the past two years has essentially come from “bigger models + longer training + stronger infrastructure,” with leading companies building a capital-intensive advantage before everyone else. But the counterargument also holds: more parameters do not automatically mean a better product. Data quality, training methods, inference efficiency, and safety alignment—the full set of methods used to make models more stable and controllable—also matter a great deal. More importantly, most of the numbers in the post are rumors and should not be treated as confirmed facts.
Impact on regular people
For enterprise IT: if leadership mainly comes from scale, the cost-performance case for building a top-tier foundation model in-house gets even weaker, and buying and integrating external models may be the more realistic path.
For individual careers: we should pay more attention to who actually plugs models into workflows, rather than staring only at parameter leaderboards; what really changes productivity is often the way deployment is done.
For the consumer market: big companies will keep spending heavily to build stronger models, while smaller companies will likely look for openings in price, speed, and vertical use cases. What users will see is a product market with more visible tiering.