Copyright © HT Digital Streams Limit all rights reserved. Christopher Mims, The Wall Street Journal 6 min Read 30 Aug 2025, 03:13 PM IST Illustration: Daniel Hertzberg Summary With models that think more than ever, the small businesses that buy artificial intelligence of the giant to create apps and services feel the pinch. As artificial intelligence became smarter, it was supposed to get too cheap for meters. It is anything but. Developers who buy AI at the barrel, for programs that do things like making software or analyze documents, discover that their accounts are higher than expected – and grow. What is the cost of cost? The latest AI models do more ‘thinking’, especially when used for deep research, AI agents and coding. Thus, while the price of a unit of AI, known as a sign, continues to decline, the number of signs needed to perform many tasks is skyrocketed. This is the opposite of what many analysts and experts even predicted a few months ago. It has a new debate in the technical world about who will be the AI winners and losers. “The arms race for whom the smartest thing can make has led to a race for whom the most expensive thing can make,” says Theo Browne, CEO of T3 Chat. Browne needs to know. His service enables people to access dozens of different AI models in one place. He can, over thousands of user inquiries, calculate his relative costs for the different models. Penny Wise, Pound-Foolish Remember, AI training and AI distractions are different. The training of those enormous models is still demanding more expensive processing, delivered by the AI super computers you probably heard. But getting answers from existing models – in the arrest – needs to get faster. Sure enough, the cost of distraction drops by a factor of 10 each year, says Ben Cottier, a former AI engineer who is now a researcher at Epoch AI, a non-profit research organization that has received financing from Openai in the past. Look at full image graph: Despite the decline in cost per sign, which increases the cost for many AI applications, so-called arguments. Many new forms of AI have re -questions to check their answers, and fan on the web to collect extra intel, even write their own small programs to calculate things, all before returning with an answer that can be as short as a sentence. And AI agents will perform a long series of actions based on user directions, which may take minutes or even hours. As a result, they make significantly better reactions, but they can spend much more signs in the process. If you give them a difficult problem, they can just keep going until they get the answer, or not try. Here are approximate amounts of tokens required for tasks at different levels based on a variety of sources: • Basic chatbot Q&A: 50 to 500 tokens • Summary of short document: 200 to 6,000 tokens • Basic Code Aid: 500 to 2000 tokens • Writing Complex Code: 20,000 to 100,000 tokens • Legal document: 75,000 to 250,000+ tokens Multi-Step tool: 100,000 to 250,000+ tokens • Multi-Step. Tokens Hence the debate: If new AI systems that use more sizes, just to answer a single request, drive a large part of the increase in the demand for AI infrastructure, who will eventually hold the bill? Ivan Zhao, CEO of the productivity software business, says his business had margins of about 90%two years ago, typical of cloud-based software businesses. Now, about 10 percentage points of that profit, are going to the AI businesses that support the latest offer of the idea. The challenges are similar – but potentially worse – for businesses that AI uses to write code for developers. These “vibecoing” businesses, including wiser and reply, have recently adjusted their prices. Some users of the cursor, under the new plan, burned through a month’s credits within a few days. As a result, some have complained or switched to competitors. Look at the full image Amjad Masad is the founder and CEO of Reprit, an AI service that helps coders and non-codeers help create programs. Photo: Bryan Bedder for the Wall Street Journal and when replit has updated its pricing model with something that calls it ‘effort-based prices’ in which more complicated requests can cost more, the world’s complaint, Reddit, filled with posts by users who declare they abandon the vibecoding app. Despite protests of a noisy minority of users, “we saw no significant income or slowdown in revenue after the pricing model was updated,” said Replit CEO Amjad Masad. The company’s business customer plan can still recommend margins from 80% to 90%, he adds. Some consolidation in the AI industry is inevitable. Warm markets eventually delay, says Martin Casado, a general partner at enterprise capital firm Andreessen Horowitz. But the fact that some AI startups sacrifice profits in the short term to expand their customers bases is not proof that they are in danger, he adds. Casado sits on the boards of several of the AI businesses that are now burning investors to expand quickly, including the cursor. He says some of the companies he puts on the boards are already striving for healthy margins. For others, it makes sense to “go for distribution just,” he adds. Myser did not respond to requests for comment. One solution: Dumber AI The large companies that create the latest AI models can at least afford to spent more than $ 100 billion a year jointly expanding infrastructure to train and deliver AI. It includes well-funded startups opening and anthropic, as well as companies such as Google and Meta that divert profits in other operations to their AI businesses. For all the investments to pay off, businesses and individuals will eventually have to spend large on these AI-powered services and products. There is an alternative, says Browne: Consumers can only use cheaper, less powerful models that need fewer resources. For T3, its very model Ai-Chatbot, Browne begins to investigate ways to encourage this behavior. Most consumers use AI chat bots for things that do not require the most resource-intensive models, and can be hampered to ‘dumber’ ais, he says. It fits the profile of the average chatgpt user. Openai CFO said in October that three -quarters of the company’s revenue from ordinary Joes and Janes pays $ 20 a month. This means that only a quarter of the revenue from the business comes from businesses and startups that pay to use its models in their own processes and products. And the price difference between well-enough AI and the latest AI is not small. The cheapest AI models, including Openai’s new GPT-5-nano, now cost about 10 cents per million signs. Compare this to Openai’s full-fledged GPT-5, which costs about $ 3.44 per million signs, if you use a weighted average of industry for use patterns, Cottier says. While tariffs and dumber ai can help some of these AI-used startups for a while, it puts them in a binder. Price increases will drive customers away. And the very big players, who own their own monster models, can lose money while serving their customers directly. At the end of June, Google offered its own code writing instrument to developers, completely free. What raises a thorny question about the state of the AI boom: How long can it last if the giants compete with their own clients? Write to Christopher MIMS on Christopher.mims@wsj.com Catch all the business news, market news, news reports and latest news updates on live currency. Download the Mint News app to get daily market updates. More Topics #Kartic Intelligence #ChatGPT Read next story
The latest AI was supposed to get cheaper. It’s more expensive than ever.
