Why LLMs Have Hit a Plateau, and What That Means for Software Development

This article does not really fit here, because it is not directly about agile work. But the topic is important enough that I am writing it down here anyway, because it concerns everyone who develops software or has software developed.

TL;DR

My thesis: In the development of large language models, we have reached a plateau. By that I do not mean that the technology will disappear again. I mean that the economic value of LLMs for serious software development is much smaller, and arrives much more slowly, than the invested capital assumes. The gain in capability since the end of 2024 came mainly from better tooling and from a form of thinking (reasoning models) that is bought with more compute, and hardly at all from fundamentally smarter models. The price per request is subsidized today and has to move toward the real cost. The measurable productivity gains show up at the level of individual tasks, but not at the level of whole companies. And the capital market is beginning to price exactly that in. Once companies have realized the real, modest efficiency gains, the pressure to invest falls, and with it the pace of further development.

Whoever reads the rest gets the derivation, with sources for the hard numbers. Whoever does not follow me in the end should please tell me at which point the chain breaks.

Two camps

In the debate about LLMs there are roughly two positions. One camp says that within a few years you will barely need finished software, because you simply have the model write whatever you happen to need. Models and the tools around them would keep getting better, and the pace would stay high. The other camp says the models themselves are barely getting smarter. What improves are the tools ("harness") and the way a model uses them, but not the actual thinking. I belong to the second camp, and this text explains why that follows logically from my point of view.

What has really improved since 2022

Between the arrival of ChatGPT at the end of 2022 and the end of 2024, the models clearly got smarter. I do not dispute that. After that, the pattern changes. Since then the noticeable progress comes less from the model itself and more from what happens around the model. An assistant that can read files, run code and operate tools appears markedly more capable, without the underlying model understanding anything fundamentally more. This environment is the real lever of the recent period.

The most common objection is that the new reasoning models are surely a genuine leap. That is half true, and the other half is decisive for my thesis. Reasoning models are partly real model progress, because thinking through intermediate steps is built into training. But they are partly just more compute at runtime, what experts call test-time compute. The model thinks longer by putting more requests to itself. This extra thinking time is not free, it is paid for in tokens. Better answers through more thinking therefore almost always mean more expensive answers. That matters, because it ties the supposed leap in capability directly to the cost question.

The prices are subsidized

Today LLMs seem cheap. That is to a good extent a market phase and not the price that has to settle in the long run. The providers are fighting for users and market share and hand out tokens below full cost. This pattern is familiar from the cloud business: at first everything was cheap, until everyone was locked in, and then the prices went up. The same is visible in the ride-hailing business, where prices rose noticeably after the subsidized early phase.

A frequent objection is that the cost per token has been falling for years rather than rising, and that is true. More efficient models and better hardware keep pushing down the cost per request. My point is a different one. The price paid today lies below the real cost, not above it, and over time it has to move toward that cost. And even where the cost per token falls, the expensive thinking of reasoning models eats part of that back up, because it consumes a multiple of the tokens. Both together argue against the idea that everything becomes smarter and practically free at the same time.

The value does not show up where it counts

The strongest objection against me is that the efficiency gains are enormous, that entire activities will soon become obsolete. If that were so, after about four years of broad LLM use one would have to see it clearly by now. You see it at the level of individual tasks, and at the level of whole companies you do not. This gap is the core of my thesis.

At the level of individual tasks the gains are real. Depending on the study, single tasks are completed faster by values in the range of roughly twenty to over fifty percent. Goldman Sachs itself cites a median of around thirty percent for individual, clearly measurable tasks. At the level of whole companies and the economy this effect almost completely disappears. The same analysis by Goldman Sachs finds no meaningful relationship between productivity and LLM adoption at the economy-wide level, and the chief economist sums up the contribution of the enormous investments to US growth in 2025 as "basically zero".

It becomes even clearer inside the companies. A 2025 MIT analysis found that around ninety-five percent of enterprise projects with generative models deliver no measurable contribution to results. In a 2026 PwC survey, a good half of the executives say they have gotten nothing out of their investments so far.

The gap shows up especially cleanly in a controlled study by METR from 2025. Experienced developers working on their own mature projects were on average nineteen percent slower with LLM tools, yet believed they had been twenty percent faster. The study is small, concerns a particular generation of tools, and METR itself now classifies it as a snapshot. But the direction matches the big numbers exactly. Perceived pace and measured result come apart, and in the daily reality of large codebases the help can even slow you down.

Vibe coding is not software development

The decisive distinction that is almost always missing from the debate is the one between two completely different activities that only happen to both produce code. Whoever mixes them up inevitably arrives at false conclusions about the market.

On one side there is what is often called vibe coding. Someone who cannot really program describes in ordinary language what they want, and the model delivers something that runs. The typical case is a small tool for one's own use, for example so that you do not have to open a presentation program for a simple task. The result only has to work now and only for this one person. Nobody has to understand the code, nobody has to fit it into a larger system, nobody maintains it over years, and nobody is liable when it stops working in half a year. The value is real, but it is private and small. For the overall picture it is almost meaningless. As soon as the requests cost real money, a large part of these makeshift tools no longer pays off anyway, and people open the presentation program again.

On the other side there is what companies like SAP or Uber actually spend money on. This software has to run reliably over years, be developed further by many people, fit into other systems, be secured, and have someone who is accountable in case of failure. Here the work consists only to a small part of writing new lines of code. The far larger part is understanding existing dependencies, reviewing, integrating, securing, maintaining and coordinating between people. These parts a model cannot take over, because they demand responsibility and an understanding of the whole that reaches beyond the individual task. A model speeds up the isolated task. It does not remove the bottleneck, and with serious software the bottleneck is almost never the typing.

This resolves the contradictory numbers. The big speedups come from the first world, from bounded individual tasks. The absent effect at the company and economy level comes from the second world, where the big money sits. Whoever concludes from a vibe-coding experience that software development is soon solved has simply confused the two worlds. Large software needs real people over years to steer it and to be accountable for it.

From this follows a sober estimate. For serious software development I consider a saving potential on the order of ten to twenty percent realistic, not the ninety percent that is sometimes claimed. Whoever names the high figure usually extrapolates a task-level speedup that does not reappear in the overall process. It is notable that even convinced proponents of the disruption thesis end up, for the orderly case, talking about roughly ten percent improvement and about rising prices once the market is divided up. That lands them at exactly the values I use here, without measuring them.

What the capital market is saying right now

There is a current case against which my thesis can be tested, because there the business was openly sold as the actual market. In 2026, xAI was merged with SpaceX and the combined company was taken public. In the offering documents, a total market of around twenty-eight trillion dollars was laid out, of which over ninety percent (26.5 of 28.5 trillion) was attributed to language models via xAI. The space share was deliberately kept small. So the story was the model business, not rockets.

The decisive number is in the same document. In 2025 xAI generated about 3.2 billion dollars in revenue, but wrote an operating loss of about 6.4 billion, which had to be covered entirely from Starlink's profit. A business sold as a future market that burns twice as much as it takes in, cross-subsidized by the rocket. That the share then fell below the offering price within a month fits the picture, even though a very small free float and the general market mood played a part. The most purely model-marketed IPO of this size rested on an astronomical market, burned money in doing so, and the market did not pay the price. That fits the suspicion that the promised market is smaller than the story.

Why this amounts to a plateau

Now the chain can be closed. The models themselves are barely getting fundamentally smarter, the progress comes from tooling and from additional compute. This compute is subsidized today and will become more expensive. The real value for serious software is closer to ten to twenty percent and barely shows up in the wider economy, because the big money sits in systems that need people. The capital market is beginning to price exactly that in.

A legitimate objection is that the large providers such as Google, Microsoft, Meta or Nvidia finance their model development out of ongoing corporate profit and are not dependent on venture capital. So the research does not stop just because the euphoria cools off. That is true, but it changes little. Those budgets, too, are governed by the expected return, and if the market turns out to be considerably smaller than promised, the willingness there also falls to buy each new model generation with ever larger and more expensive training runs. The question is not whether it continues, but at what pace. And that pace is what the first position overestimates.

Once companies have realized the modest potential, there is no longer any reason for ever new, huge investment rounds in the hope of a much larger market. But if the inflow of capital falls, so does the budget for the further development of models and tools. This slows the pace that the first position takes for granted. LLMs do not disappear, they become a useful tool with limited, well-estimable value. That is what I mean by plateau.

What would refute this thesis

So that the text stays honest, I will name the point at which it falls apart. My cost argument breaks if freely available models become so good and so efficient that they run capably on ordinary hardware locally. Then the requests would in fact be almost free, and expensive thinking would stay cheap. I do not rule out this development, and the best open models are already remarkably strong. What matters, though, is that this objection only hits the cost side. Even if the requests were free, the second part of my thesis would still hold. Large software needs real people over years to steer it sensibly and to be accountable for it. Cheap access makes the tool more accessible, but it does not turn ten to twenty percent into ninety.

Conclusion

I am not claiming that LLMs are a fad. I am claiming that the gap between the promised and the actual value for serious work is large, and that this gap will slow further development. The evidence is now in, in productivity studies, in economy-wide figures, and in the reaction of the capital market. Whoever disagrees has to show at which point the chain breaks. That is exactly what I am curious about.