Nobody drives a nail with a rocket-propelled grenade.
Anyone who is mildly good at their craft knows that efficiency beats power when you’re choosing the right tool for the job. So why are so many enterprises spending billions of dollars using the largest general-purpose systems ever built to accomplish tasks that could be handled more efficiently by small, specialized, local models?
The answer is what we call the Frontier Fallacy. People are operating under the assumption that the most capable model available is the right model for the job.
They’re wrong for a lot of reasons. But the most important is that intelligence is getting smaller.
Intelligence is getting smaller
In May 2022, the smallest model that could clear 60% on the MMLU benchmark had 540 billion parameters. Two years later, a model with fewer than four billion parameters cleared the same bar. Capability that required a data center in one product cycle fit on consumer hardware in the next.
That trend hasn’t slowed down since 2024. Just a few weeks ago, webAI open sourced TwiL-LM3, a 3 billion parameter model that beats gpt-oss-120b, a model 40 times its size, on four of five benchmarks in webAI's formal-reasoning suite.
This pattern is older than AI. Creating a new technology always takes more machinery than using it once it’s been created. When people first invented computers, they existed as mainframes that filled entire rooms. But then we created personal computers that could fit on people’s desks. Followed by smartphones.
Artificial intelligence is following the same pattern.
Even NVIDIA, one of the most important players in the mad dash to build data centers for frontier LLMs, has researchers who published a position paper arguing that small models are already powerful enough to handle many of the calls inside AI agent systems, better matched to how those systems are actually built, and cheaper to run. Where a real general-purpose conversation is needed, their recommendation is a mix: agents that draw on several different models and use the right one at each step, rather than sending everything to one large model by default.
Enterprise problems are already specific
Most enterprise problems are already well defined. Companies have sorted their business processes, built their customer channels, and accumulated decades of domain knowledge. They are not looking for a system that reinvents any of that. They want precision inside a process that already works.
A model that is decent at everything optimizes for something that businesses didn’t ask for. Zoom out far enough on a map to see all of Texas, and you can no longer find the coffee shop on Congress Avenue. The view is not wrong. It is at the wrong resolution for the question.
Why can’t we just specialize frontier models?
Defenders of cloud-based frontier LLMs are likely to point out that it’s possible to turn Claude and ChatGPT into industry specialists by curating the context we feed to the models. Assemble the right documents, index them, retrieve the relevant passages at query time, and a general-purpose model returns domain-accurate answers. But the biggest problem with this argument is economics.
Cloud models are billed by usage, so curating context makes the answers better without changing the fact that you pay for every single one of them. Picture a customer support team that finally gets this working well. The model has the product documentation, it has the past tickets, and it gives accurate answers. So the team uses it constantly. Every question from every agent on every ticket is a separate charge, and the better it works the more it gets used. Specialization succeeds and the bill goes up with it.
In other words, launching an rpg to pound in a nail isn’t just less efficient mechanically. It’s also a lot more expensive than using a hammer.
A real-world example
webAI is working with an aviation company that has a very specific challenge. So we built them a very specific solution.
Aviation mechanics and technicians often work in hangars with limited-to-no internet connectivity. They also depend on enormous technical manuals that can be tens of thousands of pages long, and they’re often required by regulations to consult those manuals before they make changes to the aircraft in order to preserve safety standards. They legally can’t rely on an LLM’s best guess at the right answer. They have to consult the original source.
In a pilot evaluation, technicians spent up to 16 hours making 37 trips back and forth between the aircraft and the terminal that held the technical manual to finish one engine change.
Between the lack of reliable internet and the regulation requirements, cloud-based frontier LLMs couldn’t help these technicians. So we built webAI Frontline. Frontline uses an industry-leading embedding model and memory compression technology to put hundreds of thousands of pages of technical documentation onto a single iPad. It lets technicians ask plain-language questions about specific procedures, then it provides the answer they’re looking for, including a primary source citation, in less than two seconds.
And because all the inference happens directly on the iPad, it works without any WiFi at all, and it doesn’t charge our customers per-token. It doesn’t cost any more for a technician to ask 100 questions than it does to ask one.
The end result was that with Frontline, those technicians reduced the time they spent searching for answers by more than 66%.
Undermining the Frontier Fallacy
webAI Frontline is just one example of smaller models providing better solutions than frontier models in enterprise AI. As intelligence keeps getting smaller, the range of work that can run locally on devices keeps expanding.
Companies who get ahead of this trend by investing in specialized, local AI will be in a far better position to deliver long-term business value than those who bet their entire AI budgets on frontier LLMs.