Dear forum members,
I wish to share a recent personal experience regarding the use of Artificial Intelligence APIs for application development, hoping it might provide useful insights to others. I’d like to begin by emphasizing how certain initial difficulties ultimately pushed me towards growth and greater awareness in choosing tools.
My foray into AI-assisted programming began a few weeks ago. With considerable enthusiasm, I started developing an application and purchased an API key for the Gemini 2.5 Pro model. However, this initial enthusiasm soon met with some challenges: the model tended to generate code affected by numerous errors and often entered redundant loops without reaching an effective solution.
Despite this, I persevered with the work for three evenings after exhausting the complimentary free credit. To my great surprise and profound regret, I discovered that the cost for using the API for that short period had climbed to an exorbitant €232. I contacted technical support, a process that was not straightforward in itself, to inquire about more advantageous pricing plans, but I was informed that no such options existed. The only suggestion I received was, in my opinion, rather impractical: to reduce the number of prompts.
This situation led me to seek alternatives. I then started using Visual Studio Code, a tool that proved to be extremely valuable. Through the CLINE extension, I initially tested the QWEN model, which, however, I did not find fully efficient for my programming needs, often having to resort to Claude for error correction.
The real breakthrough came with the purchase of an API key for Claude 3.7 Sonnet. A very positive initial aspect was the ability to pre-load credit, allowing for transparent and predictable cost control. The biggest surprise, however, came from analyzing the consumption: even using the latest 3.7 Sonnet version, the cost turned out to be surprisingly low. To provide a concrete example: I asked the model to replicate my application from a GitHub repository; the operation was performed almost entirely automatically, including connecting the implemented services, correcting previous errors attributable to Gemini, and a general qualitative improvement of the application, with minimal intervention on my part.
The total amount spent for these operations with Claude? An almost unbelievable €7. Yes, just seven euros for a significantly more productive experience. The speed and effectiveness demonstrated by the model have been, so far, decidedly satisfactory.
I felt it appropriate to share this experience to highlight how a careful evaluation of platforms and their respective cost models is fundamental. Sometimes, seemingly cutting-edge solutions can lead to significant expenses for results that are not always optimal.
I will soon be settling the amount due for the use of the Gemini API and, consequently, I will discontinue my subscription to that service.
I hope this account may be of some use to the community.
Sincerely,
R.P.
P.S. I want to make it clear: I have nothing against Google or “Firebase Studio” personally. But the problems I encountered are numerous, and the costs are simply insane! The competition, quite frankly, is just disarmingly good – especially when it comes to value and efficiency.
