Watching the AI market lately, I often get a strange feeling. Frontier models are still powerful. But I no longer see that power the way I once did. Not because their performance has declined. Because I increasingly doubt whether that performance can keep commanding a high enough price. I can imagine the order defined by OpenAI and Anthropic—the handful of companies that burned through the most capital and GPUs over the past few years to set the terms of this era—being shaken sooner than we think.
Until now, the market has treated this order almost as a law of nature. The most advanced model would have the greatest influence, and that influence would lead to the highest prices and the largest valuations. The power of frontier models rested not just on performance itself, but on the conviction that this performance could justify a high price.
What is cracking now, though, is less the technology than the way it makes money. These companies still lead on the hardest problems. But many users have quietly begun asking for enough performance rather than the best performance. In the early days of an industry, there was little choice, even if the better product cost much more. Once it becomes widely available, the question changes. What matters is less how exceptional it is than how cheaply it can be good enough.
This shift is especially clear among everyday users rather than developers. If all you need is to polish an email, summarize a report, or get answers to everyday questions, world-class reasoning is not essential. Speed matters here, too, not just performance. An advanced model may produce a more refined answer, but the slower its response and the longer it goes on, the more it becomes a drawback for ordinary users.
For many people, what matters is not the deepest answer, but a good enough answer they can use right now. Frontier models may still wear the crown, but if people do not need a king every day, the value of that crown changes.
The numbers put this feeling in starker terms. In LangChain’s Deep Agents evaluation, Claude Opus 4.6 scored 0.68 in accuracy, while GLM-5 scored 0.64. A gap remains, but it is too narrow to read as an unassailable moat. The cost gap, meanwhile, is much larger. Opus 4.6 costs $5 per million input tokens and $25 per million output tokens. GLM-5 costs $0.95 and $3.15, and MiniMax M2.7 costs around $0.30 and $1.20. LangChain noted that at 10 million output tokens a day, Opus 4.6 costs about $250, while MiniMax M2.7 costs about $12. That is a difference of roughly $87,000 a year. The crown is still there, but the market is now calculating tolls rather than admiring crowns.
The more decisive change is that the role of this advantage is itself shifting. For ordinary tasks, routing that allocates models intelligently is becoming more important than having the smartest model. Requests that a cheap model can handle are sent through as cheaply as possible, while only the truly difficult ones go to frontier models. As this structure spreads, frontier models stop being the default and become exceptions reserved for difficult tasks. The easy requests leave, and only the difficult, expensive ones remain. They are pushed into a position where they carry only the most costly work. The throne is not being vacated. It is becoming an increasingly loss-making seat.
The frontier’s crisis therefore begins with a crack in its justification, before any technical disadvantage. OpenAI and Anthropic drew their power not simply from having good models. It came from the belief that only a few giant labs could build top-tier models, and that they therefore deserved high prices and high valuations. The cracks begin not in performance charts, but when their exclusive hold on that belief starts to weaken. Open models may not have taken the crown, but they have begun to unsettle its price tag.
In this light, Google’s release of Gemma 4 as an open model is quite symbolic. It is a difficult choice for a frontier lab whose fate rests entirely on token revenue. Put another way, Google has other sources of revenue and control over distribution that allow it to survive even if it releases its models openly. The more frontier labs cling to tolls on the models themselves, the more Google’s choice looks like a sign of room to spare. Some have to sell models to survive. Others can survive even while giving them away. This difference runs deeper than it seems.
The warning signs at OpenAI are clear. Sora once symbolized OpenAI’s ambition to dominate not just text but video as well. Recently, however, it has moved toward effectively shutting it down, and the story of the Disney partnership tied to it has also begun to unravel. The market is returning to a familiar proposition: being first to show a cutting-edge demo and turning that demo into lasting market dominance are entirely different things. Amazing technology does not immediately make an amazing business.
Anthropic’s warning signs come from the opposite direction. Its move to restrict use through third-party channels like OpenClaw, announced without warning, reveals how much pressure its board is putting on the company to become profitable. Twitter is indeed full of posts from people who have switched OpenClaw’s main model to an open model or a cheaper alternative. Even if this is not yet measured evidence of a mass departure, the direction is clear. The moment cost control becomes more visible than expanding influence, a company’s hold on its ecosystem can weaken faster than expected. Empires often shrink faster when they stop expanding on their own than when they face an outside invasion.
Power is slowly moving beyond the best model. Frontier labs sell smarter models, but what the market actually buys is a more predictable system. Reliability, predictable costs, ease of integration, a close fit with workflows, ease of evaluation and management. All of these are determined more by what surrounds the model than by the model itself. Ironically, the stronger models become, the more the center of value shifts to surrounding systems such as prompt design, data pipelines, internal workflows, and routing. We are moving from an age when the king ruled everything to one where power belongs to the distribution grid that calls on the king.
What is happening now is therefore closer to reclassification than intensifying competition. The market is beginning to see OpenAI and Anthropic not as software empires with unlimited margins, but as infrastructure businesses that must endure huge upfront investments and short windows to recoup them. Models become infrastructure, and infrastructure margins eventually converge. Open models that are good enough, the spread of routing, pressure to cut prices, rising inference costs, restrictions on third-party use, rock-bottom pricing from China. When these overlap, what shakes is not the growth rate but the very premise of the valuation.
More frightening is the possibility that these cracks will not stop with the two companies. OpenAI and Anthropic are now something close to benchmarks for AI capital markets. Their high valuations have raised expectations across model companies, chipmakers, data centers, cloud providers, and application services. If the market starts seriously concluding that frontier model companies’ valuations can no longer be justified by high unit prices, high market share, and high long-term margins, the shock will not stay within individual companies. The moment the way we value the companies at the very top is shaken, all the expectations stacked beneath them are recalculated at once.
The winner of the next era may not be the lab that builds the best model. It may not be whoever creates more expensive intelligence, but whoever turns intelligence into infrastructure so commonplace that we no longer notice it. Electricity did not change civilization because it was great. The world changed when electricity became so cheap and so widespread that no one called it a special technology anymore.
So are we living through the fall of an empire, or a miracle becoming ordinary? Not the moment the crown falls to the floor, but the moment it becomes as familiar as a power outlet on the wall, and no one looks up at it anymore.
