Is AI a Luxury Reserved for Big Companies?

A Costly Dependence on AI Tools

In the tech world, it is common to become accustomed to tools that initially seem free. However, it sometimes happens that these tools, which were thought to be a given, reveal their true nature: they were actually in a trial period. Once that period is over, access becomes conditional on a paid subscription. Users then face a pricing update that, while not completely excluding them, pushes them to spend more to retain the features they have grown accustomed to.

This phenomenon is not new in the software industry, but with the rise of artificial intelligence, the pace and stakes have significantly increased. Large language models, which are at the heart of many AI applications, involve high operating costs. Companies feel intense pressure to monetize these investments, and the credit-based pricing model has become the norm for navigating this context. This has created a market where basic access to these tools and significant access have become two distinct realities. Freelancers and small teams, often with limited budgets, are the most affected by this dynamic.

A Precedent with Figma

To illustrate this trend, the example of Figma is particularly telling. At the Config 2023 event, Figma introduced Dev Mode, a tool designed to facilitate the transition between designers and developers. Initially offered in a free beta version, it was quickly adopted by cross-functional teams, becoming a central element of their workflow.

However, at the end of the beta period, Dev Mode became a paid service, costing $25 per user for the Organization plan and $35 for the Enterprise plan. Simultaneously, Figma restricted free access to the Inspect panel, thereby pushing users towards the paid version. This decision sparked significant backlash from the community, which compared Figma to Adobe for its unflattering pricing practices. Some small agencies even considered alternatives like Penpot, while licensing costs for developers sometimes doubled, significantly increasing annual expenses for Figma.

The Rise of the Credit Model

The credit model has become the dominant monetization mechanism for AI in SaaS. According to data from PricingSaaS, which tracks price changes in 500 of the top SaaS companies, 79 companies had adopted a credit model by the end of 2025. This is an increase from 35 at the end of 2024, representing a 126% year-over-year growth. Figma, HubSpot, and Salesforce all made the transition during this same period.

The logic behind this model is legitimate. AI features incur real marginal costs that traditional SaaS products simply do not have. Running a large language model costs money per request in a way that, for example, adding a user to a spreadsheet does not. As models become more capable, they also become more expensive to serve: Claude Opus 4 from Anthropic, for instance, costs $15 per million input tokens and $75 per million output tokens, among the highest rates on the market. These costs must be absorbed, and measurement systems are an attempt by companies to pass them on in a structured manner while keeping the advertised subscription price unchanged.

However, understanding why credit models exist does not mean treating all implementations as equivalent. The same structure can be a tool for transparency or a mechanism for exploitation. Recent history provides instructive contrasts.

Who is Left Behind

The credit model creates an asymmetry that is easy to overlook when the discourse focuses on the total monthly cost. The real gap is not necessarily between a $20 plan and a $40 plan. It lies between users whose employers absorb the costs of AI as a business expense and those who pay out of pocket. Not to mention teams with predictable and regular workflows versus those whose usage varies with each new project.

For a freelance designer working on Figma Make across multiple client briefs, or a small agency building prototypes under tight deadlines, the anxiety related to credits is not abstract. It changes the way you work. You ration prompts. You wonder if a design iteration is worth the cost. You turn to lower-capacity models to conserve credits, even when advanced models would produce better results. The tool is technically available. The experience of using it freely is not.

This is a structural problem that deserves to be named.