One of the things they have achieved, in quotes, the LLMS is bring the programming closer to users They don’t know how to program. It is no longer necessary, again, in quotes, to know programming languages to do simple things, but it is enough to give natural language instructions to a model and let it do its magic. This is what we know as Vibe Coding And there are startups dedicated to it, see Windsurf (now owned by Openai) and Anysphere, the matrix of Cursor.
The problem is that the accounts do not come out.
How does this work. Companies such as Cursor, Windsurf and Replit have become famous for the Vibe Coding. Basically, these offer access to a programming platform that, in the background, is based on the models of large companies of AI. In other words, Anysphere charges the user a subscription for accessing cursor and, in turn, pays Google, Anthropic and Openai for accessing Gemini’s API, Claude and GPT.
Money, money, money. The problem that this business model has is that many costs do not depend on startup itself. On the one hand, AI companies are refining their new models to program better, See GPT-5. These models, in addition, are generally more expensive as we will see later, but the Vibe Coders They love them because the results achieved are better.
Thus, the startups of Vibe Coding They are in the position of having to offer the best models (and therefore, the most expensive) without raising the price of the subscription and, at the same time, competing against the companies that create these models and that also have their platforms of Vibe Coding. Let’s think, for example, Claude Code (Anthropic) or Codex (OpenAI).
In short accounts. When the product depends practically completely on third -party models, a change in the price structure can cause the startup not to face the costs. This is the case of Windsurf, a company Vibe Coding which was sold to OpenAI for 3,000 million dollars and that, according to slide from TechcrunchHe had very negative gross margins. There is an entire industry based on the idea that the prices of the models will go down, but the reality is different, at least for the moment.
Prices of the latest Anthropic models | Image: Xataka
The models are expensive. Accessing the Google, OpenAI and Anthropic models does not come out cheap. To give some data, and based on the basis that one million tokens are more or less 750,000 words, more than the three books of ‘The Lord of the Rings’:
- Opus 4.1 (Anthropic) It has a price Input/output of 15 and 75 dollars per million tokens, respectively. It is one of the most expensive models, but the most used to program thanks to its good results.
- For prompts of up to 200,000 tokens, Gemini 2.5 Pro (Google) It has a price Input/output of 1.25 and 10 dollars per million tokens.
- GPT-5 (OpenAI) It has a price Input/output of 1.25 and 10 dollars per million tokens. Openai has thrown prices with its new model, since GPT-4.1 cost 3 and 12 dollars per million tokens.
If a company bases its business model on a subscription of $ 20 per user, it is quite likely to lose money with each of them. The best proof of this is the recent Pro change prices for cursorwhich since June 16 offers $ 20 per month of use of the large models of the invoices of API rates. Until that time, users could access 500 quick responses from the best models. The background reading is that offering broad access to the best models to program AKA to Opus 4.1 does not go to accounts.
The options. The most obvious option is that the startup develops its own models to program, something that is neither easy, or fast. Moreover, these models would have to offer capabilities equal to or higher than those of OpenAI and company, companies with many more resources to iterate and improve rapidly. This would not be viable in many cases, but in the long run it would have a positive impact on costs. In that sense, Anysphere, the company behind cursor, is already working in Fusion, its own model.
The second is to entrust to the idea that the price of AI models will go down over time, something that, for the moment, is not happening in general. OpenAi has taken steps forward With the launch of GPT-5whose price, as we saw before, has been reduced considerably. However, the trend is rather the opposite. Gemini 2.5 Pro and Claude 4.1 They are, without a doubt, models more expensive than their predecessors. The only exception, for the moment, is GPT-5 and we cannot do anything other than expect to see how prices evolve.
Cover image | THISISENGINEING
GIPHY App Key not set. Please check settings