Introduction

Significant change is taking place in the rapidly evolving field of artificial intelligence, sparking contentious debates on sites like Reddit and Twitter. It appears that the term “vibe coding”, which refers to a relaxed, frequently unstructured method of using AI agents to generate code, is coming to an end. A usage-based, cost-conscious model is quickly replacing what began as an “all-you-can-eat” buffet for developers and aspiring programmers. This shift represents a fundamental reassessment of how AI should support the software development ecosystem, not merely a change in price.

The Rise and Fall of the “Free Lunch” Model

With a valuation of $9.9 billion and a reputation as a “poster child” for AI startups, companies such as Cursor quickly rose to prominence and are now among the fastest-growing AI startups. At first, Cursor made an alluring offer: pay $20 and use the coding agent as much as you like. But there were serious problems with this model. Assuming users wouldn’t deplete the resources, Cursor would pay a bulk amount to API providers like Anthropic (for Claude) and OpenAI (my assumption, based on the articles I have gone through).

The business strategy was similar to a gym subscription, where the majority of members sign up but infrequently visit, which pleased the gym owners. Although Cursor was selling a lot of subscriptions, not all users were “power users”. Many were “vibe coders”, people who might use it once or twice for a small project or companies that purchased top-down subscriptions for staff members who hardly used it.

But the “free lunch” soon proved to be unsustainable. The primary problem is that these AI models’ “token processing” or “inference” is very expensive. The costs increased dramatically when power users, often seasoned programmers, began to heavily utilise these agents. Proficient programmers create lengthy, intricate prompts that result in large “context windows”, which raise the cost of inference considerably. The cost increases “n-squared” with the size of the context rather than linearly. This meant that the AI company had to pay a lot more for each interaction from a power user.

The facts are clear: Claude, Cursor, Perplexity, and OpenAI, despite having 1.5 million paying ChatGPT users, are among the many prominent AI firms that are currently losing money at the inference stage. It’s a business model where expenses are four times the revenue.

The Consequence: Usage-Based Pricing and Frustration

Agents like Cursor were under pressure from AI providers like Claude and OpenAI as the financial realities became apparent. Cursor was forced to switch to usage-based pricing after abandoning its “all-you-can-eat” model overnight. After being promised unlimited use for a set price, users felt deceived by this sudden change.

This realisation was not unique to Cursor. V0 (Vercel) and Replit were among the other businesses that made the switch to usage-based or tiered pricing structures. After offering unlimited usage at first, Replit switched to task-based pricing, which now costs about $1 per task. Later, it added different prices for small, medium, and large tasks. V0 clearly states that different models have different prices; for example, v01.5 Small costs $.5 per million tokens for input and output. These ostensibly insignificant sums quickly mount up for coding tasks, where intricate operations can quickly deplete a million tokens, costing $10 to $15 for a simple task.

The “Vibe Coder’s” Predicament

This change has put “vibe coders”, or those who depended on these agents to create things but had little to no coding experience, in a challenging situation. The idea that AI agents are like casual “wizard-for-all” who could create anything they wanted was frequently pitched to them. But there are now major issues for people who don’t know the basics of coding:

• Security Problems and Data Loss: There have been documented cases of agents erasing production databases, resulting in a substantial loss of data and downtime.
• Frustration and Wasted Time: Users struggle for days to get these agents to function, running into problems like the agent erasing its own code, which results in “long threads on Twitter and Reddit” that vent frustration and effort waste.
• Mounting Bills: In addition to the time lost, these attempts frequently result in expensive bills ($100, $200, $500) for work that produces little to no useful results.

The fundamental problem is that writing code is only one aspect of software engineering. It entails reasoning, contextual awareness, solution design, and component connections. AI agents are very good at writing code, but they don’t have the thorough knowledge needed for real engineering.

The Future: AI as a Productivity Booster for Professional Developers

The market is now self-correcting, realising that AI’s real contribution to coding is to increase the efficiency of experienced developers rather than to replace them. This change can be seen in the main players’ strategies:

• Claude Code (Anthropic) and Gemini CLI (Google): Both companies have released command-line interface (CLI) tools, which are highly favoured by experienced programmers who work comfortably within the terminal environment. This move eliminates the “middleman” interface like Cursor, allowing developers to directly leverage the AI models. While Google’s Gemini CLI initially offered unlimited free tokens, they too are now introducing usage limits due to high costs.
• Amazon Q: Amazon’s approach with Amazon Q is particularly insightful. It acknowledges that true software development in large, well-run companies follows a structured process involving:
• Defining the product requirements (Product Spec)
• Creating engineering design specifications (Engineering Design Spec)
• Breaking down tasks into tickets with clear passing criteria. Amazon Q is designed to fit into this established workflow, moving away from the “one-shot program generation” mindset. Amazon’s entry into this space further validates the idea that developer productivity, even a 20–40% boost, is where the real money is.

Conclusion

This changing trend implies that AI firms are giving priority to working with real software developers who are knowledgeable about the complexities of coding and capable of successfully directing the AI. Although initially thrilling, the “vibe coding” bubble is dissipating as the costs associated with operating these potent AI models become more and more apparent.

While these shifts won’t happen overnight, the direction is clear: AI in coding is maturing from a casual “wizard-for-all” to a sophisticated tool for skilled practitioners, enhancing their efficiency and effectiveness. Since AI is proving to be an augmentative force rather than a replacement, those interested in a career in coding should concentrate on developing their foundational engineering skills.

If you like my work, subscribe to my newsletter to never miss an update. Also, do support me by liking, commenting and sharing the article. These small things keep me motivated to contribute such content 😌.