The deck of punched data cards
“The cruelest prank on campus,” said dad as a brief smirk crossed his face, “was to knock someone’s stack of data cards to the floor and watch ’em scatter.” My father, Eduardo, was an industrial engineering student at the Pontificia Universidad Javeriana de Bogotá, back in the 1970s. To run code on the IBM computers of the day, students fed in precisely ordered stacks of cardboard cards with punched holes, for mathematical calculations.
If your cards were numbered, you had a way to reconstruct the deck. If they weren’t, good luck: back to square zero. Programmers punched sequence numbers into columns 73 to 80 of each card so a sorting machine could put a deck back in order, but freshmen weren’t that meticulous. I still think about that story whenever we both talk about how far we’ve come with tech. Students today have access to coding agents on their laptops that can tackle problems my father and his classmates would only have dreamt of, tokens are our new set of punched data cards
Closer than you think
I was born and raised in Colombia and spent six months in New Orleans, attending High School in the 90s. Some classmates asked whether we lived in huts, up in trees, or knew what an airplane was. I understood where their curiosity without malice was coming from: distances both geographical and technological were a real gap.
The internet has bridged that. I’m sitting at a café in “El Corralito de Piedra” (Cartagena de Indias), working with my team thousands of miles away in the US and Europe. A student in Buenos Aires, an engineer in São Paulo, a developer in Austin, they have access to much of the same technology. Still, having access to a tool isn’t quite the same as having control over it.
Someone can still knock your cards down
This time it won’t be the prankster knocking your stack down, but the cost of tokens and the ever-changing terms of service of the companies and services you depend on. Cursor’s current Plus plan includes only Grok models; third-party models on higher plans are billed separately by usage. Meanwhile, OpenAI proposed November 12 as the end of the agreement under which it supplies its models to Cursor. On October 29, Cursor reported a malicious extension posing as Claude Code in its marketplace; it was available for about 11 hours.
These are just examples of how vulnerable developers are while building workflows around tools whose pricing, availability, model selection, etc., depend on others. A pricing change or provider disagreement won’t break your code, but changes how you build it, if you can afford it and the moat built around it that impedes change.
Earning in pesos or reais, paying in U.S. dollars
For Latin American developers it’s difficult to budget given the dollar’s fluctuation. A monthly token bill denominated in U.S. dollars doesn’t stay the same in pesos or reais. Even as providers leave service prices unchanged, exchange-rate movements make your margins unpredictable. In Colombia, the dollar went from 4,478 pesos in November 2024 to 3,048 in August 2026, a 32% swing in under two years. In Brazil, the dollar jumped from 5.04 to 5.50 reais in a single day this past October. In Mexico, it climbed from 16.86 to 18.79 pesos between September 4 and October 30, 2026. Token-based billing adds another moving part to an already unpredictable budget. Gartner expects AI coding costs to pass the average developer’s salary by 2028 as token consumption grows. Choosing a model isn’t about benchmark scores only, it’s a decision about what you can afford to keep using in the upcoming months.
This isn’t a purely Latin American concern. In Stack Overflow’s 2026 Developer Survey, 32.6% of the 1,033 developers who answered a question about how cost or usage limits affect their AI use at work said they personally monitor usage, and 25.3% said they choose cheaper models when possible. Those figures come from a global, self-selected sample, not a Latin American survey. Still, they make the point: cost is already changing developer behavior.
Talent is there
LatAm is bigger than we credit it for in the developer world. GitHub’s Innovation Graph counted roughly 20.4 million developer accounts across 21 Latin American economies (Q2 2026), more than China’s 11.8 million. That includes 3.1 million in Mexico and about 1.8 million each in Colombia and Argentina. Those are GitHub accounts, not a census of employed engineers.Brazil had 10.2 million developers on GitHub as of September 2026 (GitHub’s Octoverse 2026). It added 3.28 million this year, the third most of any country.
Latin America’s developers are already using agents to write code, debug, and open pull requests, and they’re learning about the next generation of tools. Whether they can make independent choices regarding the tools they increasingly depend on is a different question.
Keep your options open
Kilo Desktop by Anaconda lets you work in a desktop workspace and choose over 500 hosted models through the Kilo Gateway, your own provider API keys, or compatible local models on your machine. It also brings code, files, terminals, git, and notebooks into the same working environment. A model dropdown lists the models your vendor has a deal with today, and those deals can change. We’ve made the case for model freedom before.
Kilo’s individual platform is free to start with. Free model options exist, paid inference can be used on a consumption basis and you can also buy Kilo credits. Local inference avoids hosted token charges but demands proper hardware. Teams on a ChatGPT plan can sign in with ChatGPT and use the usage that plan includes, rather than paying per token through the API. Auto-Efficient sends each request to the lowest-cost model Kilo has measured as capable of that kind of task. None of this makes the economics disappear. It gives you more ways to manage them.
Developers are figuring out clever ways to stretch every token. At All Day AI, some pointed out how to route simple decisions to smaller and cheaper models while using frontier ones for more elaborate work. According to OpenRouter’s Chris Clark, open-weight models now carry about 64% of the tokens but only 28% of the spending. Here is a summary of 5 best practices.
Choosing between models has its pros and cons, different models, different answers. You still have to test the output, understand the trade-offs, check how a provider handles your data, etc. Being able to make that call enables flexibility and adaptability.
Number your cards
My father’s generation learned to number their cards because they understood how fast things could become a mess. This generation needs to understand what tools we depend on, and to keep options open.
Is your company using and paying for Cursor? Kilo Enterprise is free for the rest of that contract, up to a year, if you claim it by November 30. See the offer.
Download Kilo Desktop and try the same real coding task with two models. Heck, run one of the free ones, or download a small model and run it locally. Compare the results and the cost, and see what works for you. You shouldn’t have to commit to one provider just to keep building.
There’s always a practical joker willing to knock your cards down. It’s up to you to be able to pick them up and carry on.