Skip to content

LLM ready

/en/pay-to-win-in-the-ai-bubble.md

Title:
Pay-to-win in the AI bubble: getting in is free, staying relevant is not

In 2016, anyone starting out with web, PHP and JavaScript worked and studied almost for free. VS Code had just hit version 1.0 that April, and Visual Studio Community was free for open source projects, education and small teams. A public repository on GitHub never cost anything.

To put a project online, Heroku had a Free plan, and a VPS at DigitalOcean was US$ 5 a month, close to R$ 18.50 at a time when the minimum wage was R$ 880. No subscription put one beginner ahead of another.

Back in 2016 I hardly paid for anything myself: no host, no IDE, no subscription to do my job. I started my career on Eclipse, NetBeans, Notepad++ and Sublime Text 2. There was a time when I paid for a Hostinger plan of about R$ 49, and another when I tried DigitalOcean to study, but those were optional, and nobody needed them to work on a project or land a job.

Today the free plan exists, and you can get in without paying. What costs money is what comes after the way in: staying competitive, and who decides that you need it.

The free plan lets you in

You can start today without paying anything. GitHub Copilot has a Free plan, the Gemini CLI gives you 1,000 requests a day with a Google account (according to the docs, updated on 18/06/2026) and Ollama charges US$ 0 on pay-as-you-go, on top of the models that run locally. Anthropic's free plan, on the other hand, does not include Claude Code.

Nobody is banned from starting. The difference shows up for the people who stay on free: a low limit, and no top model.

The next tier up costs dollars and comes rationed

Vendors sell tiers at US$ 20 and US$ 100, and most have one at US$ 200; OpenAI also has one at US$ 500. These are the prices on 19/11/2026, before tax:

Vendor US$ 100 tier US$ 200 tier or above
Anthropic Max Ultra (US$ 200)
OpenAI Pro 100 Pro 200 and Pro 500
Cursor Super Ultra (US$ 200)
GitHub Copilot Max none
Ollama Max (for people who run several agents) none

The detail that matters is the rationing. At Anthropic, the model the site describes as the top one (Fable) is only available up to half of the weekly limit on Max and Ultra, so not even US$ 200 gets you the best model full time. At OpenAI, Ultrafast mode only exists on Pro 500. Cursor charges for third-party models (Claude, GPT, Gemini) separately, at API list price. And on 29/09 OpenAI announced a cut to the Pro 200 allowance and then canceled it, which shows the limit can change without you having done anything.

The market forces it, not the vendor

AI adoption is almost universal. In the Stack Overflow Developer Survey 2026, 65.5% of respondents used Claude Code and 58.7% used Copilot in the past year (12,255 responses). Among the Brazilian devs in the Código Fonte TV survey, 98.47% use AI at work. If everybody uses it, using it sets nobody apart. What sets you apart is how much access you have.

Mustafa Orhun Kose put it best, in a LinkedIn post that opens with "Software development has become a pay-to-win game". He separates paying for a better IDE from paying for what looks like extra engineering capacity: one dev can run several agents and iterate all day, while the other manages tokens or waits for the limit to reset.

The post, as he published it:

Screenshot of Mustafa Orhun Kose's LinkedIn post, titled "Software development has become a pay-to-win game", in English, which contrasts paying for a better IDE with paying for extra engineering capacity and ends by saying that two devs of similar skill can have very different productivity ceilings, just because of what they can pay for.

Post by Mustafa Orhun Kose, on LinkedIn.

For people just starting out, this has a practical consequence. To show and keep your experience, you need to know how to use the top models and the agent workflows that companies charge for, and those models and limits only come with paid plans. I say this as someone living in the market as a senior programmer: from what I see in the job openings I follow, the vast majority require it so you can get ahead of the other candidates, and the people competing with you are already paying. Nobody forces you to pay to get in. The market forces you to pay to stay relevant.

A Hacker News user says he has talented classmates on free Copilot only, who are "likely being screened out in favour of rich classmates with $200/month Claude subs". In the same discussion, a reply pointed out that students usually get access through their university, and that US$ 2,000 a month is still cheaper than college.

The subsidy does not change the math

The subscription is subsidized, according to several accounts. Simon Willison says he burns the equivalent of about US$ 1,000 a month in tokens with each of two vendors, paying US$ 100 per vendor. LeadDev ran the numbers: US$ 2,180 at per-token prices against US$ 200 paid. And Fabio Akita, who uses this every day, wrote on 24/05/2026: "Paying for tokens directly is expensive. Paying for a subsidized monthly plan comes out much cheaper."

What happens inside companies shows that the spending ceiling varies a lot. According to the Pragmatic Engineer, a quantitative fund reached an average of US$ 1,600 per dev per month, with a cap of US$ 25,000 per employee, while a publicly traded company with 1,500 engineers caps it at US$ 150. The same article says company size and AI spend have no relationship.

On where prices go from here, Karol Attekita, in the video The AI bill has arrived. And the dev of the future pays it, from 19/11/2026, explains that for the numbers to work for the whole chain (the tool, the model and the infrastructure), AI should cost a lot more for whoever consumes it at the end, and that this is not happening. To get to a margin, according to her, companies "are going to have to raise prices and the bill will weigh not only on tech companies, but on developers too". She also notes that people who became dependent on the tool would be willing to pay any price to keep using it.

If the subsidy ends, the bill goes up for everybody. But the argument holds even with the subsidy: with prices below cost, the people who can pay already study and practice with tools that the people who cannot pay never even try.

2016 had paid stuff too

In 2016 there was already the Apple developer account (US$ 99 a year), private GitHub (US$ 7 a month, unlimited since 11/05/2016), IntelliJ IDEA Ultimate (US$ 149 the first year) and Visual Studio Professional with MSDN, at US$ 1,199 for a new license and US$ 799 for the renewal. Almost all of it had a free alternative: VS Code, Visual Studio Community, public GitHub, Heroku Free. Whoever paid for a better IDE or for a server to host a study project did it by choice and as a luxury, not to get ahead of other developers in the job market.

You can measure how much the optional top end cost. The math uses the average 2016 dollar (R$ 3.4833, according to Captário) and the 6.38% IOF (Brazil's tax on foreign card purchases) on cards abroad, and today's (R$ 5.888 on 19/11/2026, per InfoMoney, with 3.5% IOF). In 2016:

  • A US$ 10 VPS, GitHub, IntelliJ Ultimate (US$ 149 divided by 12) and the Apple account (US$ 99 divided by 12) add up to about US$ 38 a month, close to R$ 140, or 16% of the minimum wage.
  • On the Microsoft track (VPS, GitHub and Visual Studio Professional at the renewal price), it comes to US$ 84 a month, about R$ 310, or 35% of the minimum.

In 2026, the US$ 200 plan costs about R$ 1,219, which is 75% of the R$ 1,621 minimum wage. Add a VPS and it goes to 79%.

The exchange rate does not explain it: the dollar rose 69% from 2016 to today, and the minimum wage rose 84%. What weighs is the dollar price of the top end, which went from US$ 38 to US$ 84 a month up to US$ 200 to US$ 500. And the nature of it changed. A US$ 149 IDE did not write code for you; the token is work capacity, recurring and proportional to what you deliver.

These numbers are my own estimates from each source's prices, without the bank's margin, and they compare different things (a tool against capacity). Treat them as an order of magnitude.

In Brazil, the bill comes in dollars

OpenAI and Anthropic do not have prices in reais; billing is in dollars, changes month to month and IOF goes on top. In September, with the dollar at R$ 5.40, Exame calculated that the US$ 200 plan was 69% of the minimum wage. Since then the dollar has risen about 9%. Google is the exception: it charges in reais, R$ 96.99 for AI Pro.

Against the average reported pay in the Código Fonte TV salary survey (17,046 respondents, a convenience sample with audience bias), US$ 200 is 24% of a junior's salary (R$ 5,060.34), 14.4% of a mid-level's (R$ 8,466.52) and 8.3% of a senior's (R$ 14,607.85). That is why "US$ 200 is cheap" works for people who are already at the top and does not work for the base.

Where the argument loses strength

Open models close a good part of the gap. On the Code Arena WebDev of 04/10/2026, the best open model (Kimi K3) was 4th, with 1,674 points, against 1,765 for the first, which is closed (Danube Labs). Plans of US$ 6 to US$ 10 give access to several of these models (KDnuggets). Someone paying US$ 6 can get work done. That is why pay-to-win here means "pay to get a little ahead", not "pay to exist". That is the formulation I stand behind.

Parallelism has a human ceiling, and it is the strongest counterargument against the run-several-agents part. Simon Willison says, in a loose translation, that he can start four agents in parallel and by 11 a.m. he is wiped out. The quote is in the text by Addy Osmani, who talks about 3 to 4 threads as a personal ceiling. Peter Fulop finds two comfortable and three possible if the tasks are well separated. If the ceiling is 3 or 4 agents, that fits in a plan of US$ 100 or less, which weakens "only whoever pays more gets ahead" (that is interpretation, nobody measured it).

Productivity is uncertain too. The METR study with 16 experienced devs and 246 tasks, in early 2025, found the devs 19% slower with AI, even though they thought they had become 20% faster. The page itself warns that the result is outdated. The 2026 update brought gains of 18% and 4%, both statistically inconclusive, and METR says the data is not very reliable, partly because devs who refused to work without AI dropped out of the study. I also could not find a controlled study of several agents against a single agent. What exists are reports from companies, like Ix Technologies, which according to the Pragmatic Engineer spends US$ 3,000 per dev per month and says a pair does the work of 5 to 7 people, next to companies where, in one engineer's words, speed "did not increase much".

Akita pays, uses it, and disagrees with part of this

Fabio Akita is a Brazilian dev who uses this heavily and writes about what he pays. He pays out of his own pocket (on 14/05/2026 he wrote: "I'm retired, I don't need to charge anything"). In February 2026 he said he was on the Max 5x plan, "which I think costs about USD 100, that is, about USD 6 a day" (post of 01/02/2026). In April, on Max 20x, the weekly limit hit 100% and he had extra spend above R$ 300: "For those who think generative AI becomes free once you've subscribed to the plan, the bills beg to differ" (09/04/2026).

He confirms the subsidy math. In the table of the 05/04/2026 post, for heavy users, "Max 20x at $200/month comes out cheaper than paying for Opus per token on OpenRouter (~$450/month)". And he switches between Claude Code and Codex "because the plans have limits and I work many hours in a row" (24/05/2026). In the same table, the entry tier is the one in my argument: a free model but with a usage limit, and on the paid tier, "Pro at $20/month covers light use".

He talks about this live too. In the talk of 28/05/2026, "My Experience with Agile Vibe Coding", he describes his heavy AI use and names the method: "now I call this agile vibe coding", made of testing, checking, monitoring, reviewing and throwing away the code that does not work. In it he comes back to the subsidy ("Spend all the tokens, taking advantage now that it is hyper-subsidized and hyper-cheap"), to the US$ 2,000 that he says is cheap, adding up Claude Max 20x and ElevenLabs on a translation project, and to the market that had already blown up with the layoffs of 2022.

He disagrees with part of the argument, and three of his points I had to accept:

  • On parallel agents, on 25/04/2026: "spending a weekend setting up orchestration for five agents in parallel to save $30/month in tokens, before you have validated whether the product you are building is any good" is theater.
  • On skill, on 14/05/2026: "AI does not replace a good programmer. AI mirrors and accelerates who you are." That is compatible with the argument (money without skill gives you volume, not an edge), but it takes weight off "pay and win".
  • On open models, in the 30/07/2026 ranking: Kimi K3, on the Moderato plan, tied with Claude Opus 5 (95 points each) and was 1 point behind the first.

On the 2016 versus 2026 argument, he pokes at that too. For beginners, he wrote on 16/10/2026 that "The market will reward whoever makes that extra effort" and that "Adapting includes accepting that getting in has become more competitive". That supports the idea that it is the market doing the charging. But he does not say the barrier is the subscription: he says it is real projects, reading code, studying. And he blames the tougher market on the bubble that burst with the layoffs of late 2022, "BEFORE the advent of LLMs for programming" (08/02/2026).

Cost is the point where I agree with him. On 25/04/2026 he wrote: "$20-200/month subscriptions are trivial compared to any other professional learning cost. People spend more on bad online courses or a weekend of clubbing." He speaks as a senior dev, retired and paid in dollars, without mentioning exchange rate, IOF or salary, which is why the math in reais further up matters for the base. But his sentence reinforces my point. If US$ 20 to 200 a month is trivial, paying a monthly recurring fee has become an obligation for programmers who want to work and study the AI hype. Whoever does not pay gets left behind, and the market, from what I see in the job openings I follow, demands it in most of them so you can keep up with the people who do.

What I pay today

This month my AI bill came to about R$ 509 plus US$ 25 in JEVI (TypeSafe.ai) credits, which I signed up for just now and use by credits. Claude Pro is R$ 110.00. X Premium, which includes Grok, Cursor and Grok Bot, is R$ 77.45 with a discount until 4/12/2026. OpenCode Go Plus cost US$ 40, which became R$ 224.53 on the receipt (dollar at R$ 5.6131). Google AI Pro Family is R$ 96.99: the family plan is shared with my wife, but the whole bill is mine, every month. Outside the recurring bill, in October I bought R$ 90.00 of prepaid extra Claude usage.

In each project, the agents that stayed active this month were Claude Code, Grok, Codex, Agy (the Antigravity CLI) and OpenCode. I run more than one, yes. Claude Code and Grok are the constants every month. I canceled OpenCode and I am re-evaluating Codex, Antigravity (which comes with Google AI Pro) and JEVI, to see whether it stays next month.

In December the recurring bill drops to R$ 361.89: Claude Pro (R$ 110.00), X Premium at R$ 154.90 and Google AI Pro Family (R$ 96.99), already without OpenCode. JEVI credits stay out of that bill, because I do not know yet whether I will keep it.

When the limit runs out, I sit idle

When I hit the 5-hour or weekly limit on my agents, I sit idle. I step away from the terminal and go read a bit, watch a video, listen to a podcast or do things humans do: eat or look after my hobbies. It is the rationing of the next tier up showing up in my routine.

If the subsidy ended, I do not know what I would do. I have not worked out a strategy for it, and that is my point: when it gets even more expensive, the only people who will get ahead are those who can afford to pay or have the privilege of working at a company that pays for AI so the dev can work and study. The rest get left behind.

429 Too Many Requests

---
title: Pay-to-win in the AI bubble: getting in is free, staying relevant is not
url: https://nandomoreira.dev/en/pay-to-win-in-the-ai-bubble
date: 2026-11-19
---

In 2016, anyone starting out with web, PHP and JavaScript worked and studied almost for free. [VS Code](https://code.visualstudio.com/blogs/2016/04/14/vscode-1.0) had just hit version 1.0 that April, and [Visual Studio Community](https://www.informationweek.com/it-leadership/microsoft-streamlines-visual-studio-pricing-structure) was free for open source projects, education and small teams. A public repository on GitHub never cost anything.

To put a project online, [Heroku](https://www.heroku.com/blog/dynos-pricing-ga/) had a Free plan, and a VPS at [DigitalOcean](https://web.archive.org/web/20160208230818/www.digitalocean.com/?refcode=42241824138c) was US$ 5 a month, close to R$ 18.50 at a time when the minimum wage was [R$ 880](http://www.planalto.gov.br/ccivil_03/_ato2015-2018/2015/decreto/d8618.htm). No subscription put one beginner ahead of another.

Back in 2016 I hardly paid for anything myself: no host, no IDE, no subscription to do my job. I started my career on Eclipse, NetBeans, Notepad++ and Sublime Text 2. There was a time when I paid for a Hostinger plan of about R$ 49, and another when I tried DigitalOcean to study, but those were optional, and nobody needed them to work on a project or land a job.

Today the free plan exists, and you can get in without paying. What costs money is what comes after the way in: staying competitive, and who decides that you need it.

## The free plan lets you in

You can start today without paying anything. [GitHub Copilot](https://github.com/features/copilot/plans) has a Free plan, the [Gemini CLI](https://geminicli.com/docs/resources/quota-and-pricing/) gives you 1,000 requests a day with a Google account (according to the docs, updated on 18/06/2026) and [Ollama](https://ollama.com/pricing) charges US$ 0 on pay-as-you-go, on top of the models that run locally. Anthropic's free plan, on the other hand, [does not include Claude Code](https://claude.com/pricing).

Nobody is banned from starting. The difference shows up for the people who stay on free: a low limit, and no top model.

## The next tier up costs dollars and comes rationed

Vendors sell tiers at US$ 20 and US$ 100, and most have one at US$ 200; OpenAI also has one at US$ 500. These are the prices on 19/11/2026, before tax:

| Vendor | US$ 100 tier | US$ 200 tier or above |
|---|---|---|
| [Anthropic](https://claude.com/pricing) | Max | Ultra (US$ 200) |
| [OpenAI](https://help.openai.com/en/articles/9793128-about-chatgpt-pro-tiers) | Pro 100 | Pro 200 and Pro 500 |
| [Cursor](https://cursor.com/docs/models-and-pricing) | Super | Ultra (US$ 200) |
| [GitHub Copilot](https://github.com/features/copilot/plans) | Max | none |
| [Ollama](https://ollama.com/pricing) | Max (for people who run several agents) | none |

The detail that matters is the rationing. At Anthropic, the model the site describes as the top one (Fable) is only available up to half of the weekly limit on Max and Ultra, so not even US$ 200 gets you the best model full time. At OpenAI, Ultrafast mode only exists on Pro 500. Cursor charges for third-party models (Claude, GPT, Gemini) separately, at API list price. And on 29/09 OpenAI announced a cut to the Pro 200 allowance and then canceled it, which shows the limit can change without you having done anything.

## The market forces it, not the vendor

AI adoption is almost universal. In the [Stack Overflow Developer Survey 2026](https://survey.stackoverflow.co/2026/ai/data/ai-code-agent), 65.5% of respondents used Claude Code and 58.7% used Copilot in the past year (12,255 responses). Among the Brazilian devs in the Código Fonte TV survey, [98.47% use AI at work](https://vivendodeti.com.br/salario-dev-brasil-2026-pesquisa-codigo-fonte/). If everybody uses it, using it sets nobody apart. What sets you apart is how much access you have.

Mustafa Orhun Kose put it best, in a [LinkedIn post](https://www.linkedin.com/posts/cengineer13_software-development-has-become-a-pay-to-win-activity-7518917855778672640-ZmtG) that opens with "Software development has become a pay-to-win game". He separates paying for a better IDE from paying for what looks like extra engineering capacity: one dev can run several agents and iterate all day, while the other manages tokens or waits for the limit to reset.

The post, as he published it:

![Screenshot of Mustafa Orhun Kose's LinkedIn post, titled "Software development has become a pay-to-win game", in English, which contrasts paying for a better IDE with paying for extra engineering capacity and ends by saying that two devs of similar skill can have very different productivity ceilings, just because of what they can pay for.](../../assets/posts/2026/pay-to-win-linkedin-mustafa.webp)

*Post by [Mustafa Orhun Kose, on LinkedIn](https://www.linkedin.com/posts/cengineer13_software-development-has-become-a-pay-to-win-activity-7518917855778672640-ZmtG).*

For people just starting out, this has a practical consequence. To show and keep your experience, you need to know how to use the top models and the agent workflows that companies charge for, and those models and limits only come with paid plans. I say this as someone living in the market as a senior programmer: from what I see in the job openings I follow, the vast majority require it so you can get ahead of the other candidates, and the people competing with you are already paying. **Nobody forces you to pay to get in. The market forces you to pay to stay relevant.**

A Hacker News user says he has talented classmates [on free Copilot only](https://news.ycombinator.com/item?id=47666817), who are "likely being screened out in favour of rich classmates with $200/month Claude subs". In the same discussion, a reply pointed out that students usually get access through their university, and that US$ 2,000 a month is still cheaper than college.

## The subsidy does not change the math

The subscription is subsidized, according to several accounts. Simon Willison says he burns the equivalent of [about US$ 1,000 a month in tokens](https://simonwillison.net/2026/jun/3/uber-caps-usage/) with each of two vendors, paying US$ 100 per vendor. LeadDev ran the numbers: [US$ 2,180 at per-token prices against US$ 200 paid](https://leaddev.com/ai/cheap-ai-tokens-are-running-out). And Fabio Akita, who uses this every day, wrote on [24/05/2026](https://www.akitaonrails.com/2026/05/24/dicas-e-toolkit-de-ia-do-akita-ai-jail-ai-memory-ai-usagebar/): "Paying for tokens directly is expensive. Paying for a subsidized monthly plan comes out much cheaper."

What happens inside companies shows that the spending ceiling varies a lot. According to the [Pragmatic Engineer](https://newsletter.pragmaticengineer.com/p/whats-the-roi-of-software-engineers), a quantitative fund reached an average of US$ 1,600 per dev per month, with a cap of US$ 25,000 per employee, while a publicly traded company with 1,500 engineers caps it at US$ 150. The same article says company size and AI spend have no relationship.

On where prices go from here, Karol Attekita, in the video [The AI bill has arrived. And the dev of the future pays it](https://www.youtube.com/watch?v=SW93RhgeFHo&t=572s), from 19/11/2026, explains that for the numbers to work for the whole chain (the tool, the model and the infrastructure), AI should cost a lot more for whoever consumes it at the end, and that this is not happening. To get to a margin, according to her, companies "are going to have to raise prices and the bill will weigh not only on tech companies, but on developers too". She also notes that people who became dependent on the tool would be willing to pay any price to keep using it.

If the subsidy ends, the bill goes up for everybody. But the argument holds even with the subsidy: with prices below cost, the people who can pay already study and practice with tools that the people who cannot pay never even try.

## 2016 had paid stuff too

In 2016 there was already the [Apple developer account](https://techcrunch.com/2015/06/08/apples-new-developer-program-offers-ios-os-x-and-watchos-tools-for-99/) (US$ 99 a year), [private GitHub](https://github.blog/news-insights/product-news/introducing-unlimited-private-repositories/) (US$ 7 a month, unlimited since 11/05/2016), [IntelliJ IDEA Ultimate](https://www.infoq.com/news/2015/11/intellij-idea-15/) (US$ 149 the first year) and Visual Studio Professional with MSDN, at US$ 1,199 for a new license and US$ 799 for the renewal. Almost all of it had a free alternative: VS Code, Visual Studio Community, public GitHub, Heroku Free. Whoever paid for a better IDE or for a server to host a study project did it by choice and as a luxury, not to get ahead of other developers in the job market.

You can measure how much the optional top end cost. The math uses the average 2016 dollar (R$ 3.4833, according to [Captário](https://captario.com.br/cotacao/dolar/2016)) and the [6.38% IOF](https://www1.folha.uol.com.br/mercado/2025/05/de-fhc-a-lula-iof-de-cartao-internacional-mudou-seis-vezes-nos-ultimos-20-anos-veja-grafico.shtml) (Brazil's tax on foreign card purchases) on cards abroad, and today's (R$ 5.888 on 19/11/2026, per [InfoMoney](https://www.infomoney.com.br/mercados/dolar-hoje-abertura-fechamento-comercial-turismo-19112026/), with 3.5% IOF). In 2016:

- A US$ 10 VPS, GitHub, IntelliJ Ultimate (US$ 149 divided by 12) and the Apple account (US$ 99 divided by 12) add up to about US$ 38 a month, close to R$ 140, or 16% of the minimum wage.
- On the Microsoft track (VPS, GitHub and Visual Studio Professional at the renewal price), it comes to US$ 84 a month, about R$ 310, or 35% of the minimum.

In 2026, the US$ 200 plan costs about R$ 1,219, which is 75% of the R$ 1,621 minimum wage. Add a VPS and it goes to 79%.

The exchange rate does not explain it: the dollar rose 69% from 2016 to today, and the minimum wage rose 84%. What weighs is the dollar price of the top end, which went from US$ 38 to US$ 84 a month up to US$ 200 to US$ 500. And the nature of it changed. A US$ 149 IDE did not write code for you; the token is work capacity, recurring and proportional to what you deliver.

These numbers are my own estimates from each source's prices, without the bank's margin, and they compare different things (a tool against capacity). Treat them as an order of magnitude.

## In Brazil, the bill comes in dollars

OpenAI and Anthropic [do not have prices in reais](https://exame.com/tecnologia/examelab/quanto-do-seu-salario-vai-embora-com-uma-assinatura-de-ia-veja-a-conta/); billing is in dollars, changes month to month and IOF goes on top. In September, with the dollar at R$ 5.40, Exame calculated that the US$ 200 plan was 69% of the minimum wage. Since then the dollar has risen about 9%. Google is the exception: it charges in reais, R$ 96.99 for AI Pro.

Against the average reported pay in the [Código Fonte TV salary survey](https://pesquisa.codigofonte.com.br/2026) (17,046 respondents, a convenience sample with audience bias), US$ 200 is 24% of a junior's salary (R$ 5,060.34), 14.4% of a mid-level's (R$ 8,466.52) and 8.3% of a senior's (R$ 14,607.85). That is why "US$ 200 is cheap" works for people who are already at the top and does not work for the base.

## Where the argument loses strength

Open models close a good part of the gap. On the Code Arena WebDev of 04/10/2026, the best open model (Kimi K3) was 4th, with 1,674 points, against 1,765 for the first, which is closed ([Danube Labs](https://danubelabs.net/en/blog/best-open-weight-coding-models)). Plans of US$ 6 to US$ 10 give access to several of these models ([KDnuggets](https://www.kdnuggets.com/7-best-ai-coding-plans-for-10-or-less-in-2026)). Someone paying US$ 6 can get work done. That is why pay-to-win here means "pay to get a little ahead", not "pay to exist". That is the formulation I stand behind.

Parallelism has a human ceiling, and it is the strongest counterargument against the run-several-agents part. Simon Willison says, in a loose translation, that he can start four agents in parallel and by 11 a.m. he is wiped out. The quote is in the text by [Addy Osmani](https://addyosmani.com/blog/cognitive-parallel-agents/), who talks about 3 to 4 threads as a personal ceiling. [Peter Fulop](https://peterfulop.tech/p/parallel-agents-hit-a-human-limit-faster-than-a-system-limit/) finds two comfortable and three possible if the tasks are well separated. If the ceiling is 3 or 4 agents, that fits in a plan of US$ 100 or less, which weakens "only whoever pays more gets ahead" (that is interpretation, nobody measured it).

Productivity is uncertain too. The [METR study](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/) with 16 experienced devs and 246 tasks, in early 2025, found the devs 19% slower with AI, even though they thought they had become 20% faster. The page itself warns that the result is outdated. The [2026 update](https://metr.org/blog/2026-02-24-uplift-update/) brought gains of 18% and 4%, both statistically inconclusive, and METR says the data is not very reliable, partly because devs who refused to work without AI dropped out of the study. I also could not find a controlled study of several agents against a single agent. What exists are reports from companies, like Ix Technologies, which according to the Pragmatic Engineer spends US$ 3,000 per dev per month and says a pair does the work of 5 to 7 people, next to companies where, in one engineer's words, speed "did not increase much".

## Akita pays, uses it, and disagrees with part of this

[Fabio Akita](https://www.akitaonrails.com) is a Brazilian dev who uses this heavily and writes about what he pays. He pays out of his own pocket (on [14/05/2026](https://www.akitaonrails.com/2026/05/14/terminando-maratona-ia-sucesso-ou-fracasso/) he wrote: "I'm retired, I don't need to charge anything"). In February 2026 he said he was on the Max 5x plan, "which I think costs about USD 100, that is, about USD 6 a day" ([post of 01/02/2026](https://www.akitaonrails.com/2026/02/01/vibe-code-fiz-um-editor-de-markdown-do-zero-com-claude-code-frankmd-parte-2/)). In April, on Max 20x, the weekly limit hit 100% and he had extra spend above R$ 300: "For those who think generative AI becomes free once you've subscribed to the plan, the bills beg to differ" ([09/04/2026](https://www.akitaonrails.com/2026/04/09/como-a-elevenlabs-nao-foi-morta-pelo-qwen3-tts/)).

He confirms the subsidy math. In the table of the [05/04/2026 post](https://www.akitaonrails.com/2026/04/05/testando-llms-open-source-e-comerciais-quem-consegue-bater-o-claude-opus/), for heavy users, "Max 20x at $200/month comes out cheaper than paying for Opus per token on OpenRouter (~$450/month)". And he switches between Claude Code and Codex "because the plans have limits and I work many hours in a row" ([24/05/2026](https://www.akitaonrails.com/2026/05/24/dicas-e-toolkit-de-ia-do-akita-ai-jail-ai-memory-ai-usagebar/)). In the same table, the entry tier is the one in my argument: a free model but with a usage limit, and on the paid tier, "Pro at $20/month covers light use".

He talks about this live too. In the [talk of 28/05/2026, "My Experience with Agile Vibe Coding"](https://www.youtube.com/watch?v=U3bZavG8qQY&t=2358s), he describes his heavy AI use and names the method: "now I call this agile vibe coding", made of testing, checking, monitoring, reviewing and throwing away the code that does not work. In it he comes back to the subsidy ("Spend all the tokens, taking advantage now that it is hyper-subsidized and hyper-cheap"), to the US$ 2,000 that he says is cheap, adding up Claude Max 20x and ElevenLabs on a translation project, and to the market that had already blown up with the layoffs of 2022.

He disagrees with part of the argument, and three of his points I had to accept:

- On parallel agents, on [25/04/2026](https://www.akitaonrails.com/2026/04/25/llm-benchmarks-vale-a-pena-misturar-2-modelos/): "spending a weekend setting up orchestration for five agents in parallel to save $30/month in tokens, before you have validated whether the product you are building is any good" is theater.
- On skill, on 14/05/2026: "AI does not replace a good programmer. AI mirrors and accelerates who you are." That is compatible with the argument (money without skill gives you volume, not an edge), but it takes weight off "pay and win".
- On open models, in the [30/07/2026](https://www.akitaonrails.com/2026/07/30/novo-llm-benchmark-refiz-todos-os-testes/) ranking: Kimi K3, on the Moderato plan, tied with Claude Opus 5 (95 points each) and was 1 point behind the first.

On the 2016 versus 2026 argument, he pokes at that too. For beginners, he [wrote on 16/10/2026](https://www.akitaonrails.com/2026/10/16/estamos-no-futuro-carros-voam-e-nao-temos-mais-que-fazer-html-css/) that "The market will reward whoever makes that extra effort" and that "Adapting includes accepting that getting in has become more competitive". That supports the idea that it is the market doing the charging. But he does not say the barrier is the subscription: he says it is real projects, reading code, studying. And he blames the tougher market on the bubble that burst with the layoffs of late 2022, "BEFORE the advent of LLMs for programming" ([08/02/2026](https://www.akitaonrails.com/2026/02/08/rant-ia-acabou-com-programadores/)).

Cost is the point where I agree with him. On [25/04/2026](https://www.akitaonrails.com/2026/04/25/llm-benchmarks-vale-a-pena-misturar-2-modelos/) he wrote: "$20-200/month subscriptions are trivial compared to any other professional learning cost. People spend more on bad online courses or a weekend of clubbing." He speaks as a senior dev, retired and paid in dollars, without mentioning exchange rate, IOF or salary, which is why the math in reais further up matters for the base. But his sentence reinforces my point. If US$ 20 to 200 a month is trivial, paying a monthly recurring fee has become an obligation for programmers who want to work and study the AI hype. Whoever does not pay gets left behind, and the market, from what I see in the job openings I follow, demands it in most of them so you can keep up with the people who do.

## What I pay today

This month my AI bill came to about R$ 509 plus US$ 25 in JEVI (TypeSafe.ai) credits, which I signed up for just now and use by credits. Claude Pro is R$ 110.00. X Premium, which includes Grok, Cursor and Grok Bot, is R$ 77.45 with a discount until 4/12/2026. OpenCode Go Plus cost US$ 40, which became R$ 224.53 on the receipt (dollar at R$ 5.6131). Google AI Pro Family is R$ 96.99: the family plan is shared with my wife, but the whole bill is mine, every month. Outside the recurring bill, in October I bought R$ 90.00 of prepaid extra Claude usage.

In each project, the agents that stayed active this month were Claude Code, Grok, Codex, Agy (the Antigravity CLI) and OpenCode. I run more than one, yes. Claude Code and Grok are the constants every month. I canceled OpenCode and I am re-evaluating Codex, Antigravity (which comes with Google AI Pro) and JEVI, to see whether it stays next month.

In December the recurring bill drops to R$ 361.89: Claude Pro (R$ 110.00), X Premium at R$ 154.90 and Google AI Pro Family (R$ 96.99), already without OpenCode. JEVI credits stay out of that bill, because I do not know yet whether I will keep it.

## When the limit runs out, I sit idle

When I hit the 5-hour or weekly limit on my agents, I sit idle. I step away from the terminal and go read a bit, watch a video, listen to a podcast or do things humans do: eat or look after my hobbies. It is the rationing of the next tier up showing up in my routine.

If the subsidy ended, I do not know what I would do. I have not worked out a strategy for it, and that is my point: when it gets even more expensive, the only people who will get ahead are those who can afford to pay or have the privilege of working at a company that pays for AI so the dev can work and study. The rest get left behind.

`429 Too Many Requests`

Ready to share.