Claude Code starts at twenty dollars a month, Codex at eight, and Google's command-line agent runs a thousand model requests a day for nothing. None of those numbers are what stops people. The bank does.
Start with what costs nothing. A personal Google account gets you a free tier of a thousand model requests a day, capped at sixty a minute, which is more than most side projects burn through in an evening. Codex has its own free tier, and its Go plan runs eight dollars a month for light work. Claude includes Claude Code from the twenty-dollar Pro plan upward, or seventeen a month if you pay for a full year.
So the entry fee is small. Getting the money out of Zimbabwe and onto a foreign checkout page is the part the tool comparisons skip.
Stanbic launched a prepaid USD Visa card on May 21, 2026, aimed at exactly this problem. You load it yourself; there is no credit line and no monthly maintenance fee, and the paperwork is an ID, proof of residence, and passport photos. You do not need an existing USD account with the bank, which is the wall most first-time subscribers hit.
That changes the arithmetic more than any tool comparison does. Zimbabweans abroad sent home 2.45 billion dollars in 2025, and moving money up from South Africa can still cost as much as 12.7 percent of the transfer. With a working card in hand, the old assumption that serious tool building happens in the diaspora looks like a payments problem rather than a talent one.
That middle ground is where budgets break. You can outgrow Plus in a week of real work, and the next rung costs five times as much, which is a serious monthly commitment on local earnings. Metered access through the API is the honest answer in between, because you pay for what you use, and nothing for the days you never open the terminal.
Watch the dates as well. Sonnet 5 sits at two dollars per million input tokens and ten per million output until August 31, 2026, then moves to three and fifteen. Plan a build around today's rate, and you will be paying half as much again by September.
The flip comes when you use it daily. At twenty dollars flat, you stop counting tokens, and counting tokens while you work is its own drain on concentration.
Two discounts get ignored constantly. The batch API takes fifty percent off both input and output tokens if you can wait for the results, which suits overnight jobs like crawling a client site or scoring a few thousand pages. Prompt caching charges cache hits at a tenth of the base input rate, so the second run against a large file costs far less than the first.
Then point the cheap model at the boring work. Opus 5 runs five dollars per million input tokens and twenty-five per million output, five times what Haiku 4.5 charges on both counts, and most of what you build will never need it. Parsing a file, renaming variables, and formatting a report all run fine on the smaller model. Save the expensive one for the logic you cannot reason through yourself.
Start with what costs nothing. A personal Google account gets you a free tier of a thousand model requests a day, capped at sixty a minute, which is more than most side projects burn through in an evening. Codex has its own free tier, and its Go plan runs eight dollars a month for light work. Claude includes Claude Code from the twenty-dollar Pro plan upward, or seventeen a month if you pay for a full year.
So the entry fee is small. Getting the money out of Zimbabwe and onto a foreign checkout page is the part the tool comparisons skip.
The card is the harder half of the bill
Around seventy percent of local transactions already run in United States dollars, and most people still have no clean way to spend those dollars online. That gap is structural. It says nothing about your ability to build.Stanbic launched a prepaid USD Visa card on May 21, 2026, aimed at exactly this problem. You load it yourself; there is no credit line and no monthly maintenance fee, and the paperwork is an ID, proof of residence, and passport photos. You do not need an existing USD account with the bank, which is the wall most first-time subscribers hit.
That changes the arithmetic more than any tool comparison does. Zimbabweans abroad sent home 2.45 billion dollars in 2025, and moving money up from South Africa can still cost as much as 12.7 percent of the transfer. With a working card in hand, the old assumption that serious tool building happens in the diaspora looks like a payments problem rather than a talent one.
What twenty dollars a month does not buy
Flat plans are rationed by window, not by month. On Codex, the twenty-dollar Plus plan gives you somewhere between fifteen and eighty local messages every five hours on GPT-5.5, and the hundred-dollar Pro tier raises that to between eighty and four hundred. Claude's Max plans also start at a hundred dollars.That middle ground is where budgets break. You can outgrow Plus in a week of real work, and the next rung costs five times as much, which is a serious monthly commitment on local earnings. Metered access through the API is the honest answer in between, because you pay for what you use, and nothing for the days you never open the terminal.
Watch the dates as well. Sonnet 5 sits at two dollars per million input tokens and ten per million output until August 31, 2026, then moves to three and fifteen. Plan a build around today's rate, and you will be paying half as much again by September.
The cheapest setup that still ships work
Metered pricing beats a subscription when your work comes in bursts. Haiku 4.5 costs one dollar per million input tokens and five per million output, so a script that reads a few hundred lines and writes a few hundred back lands at a cent or two per run. Debug it twenty times in an afternoon, and you have spent under a dollar.The flip comes when you use it daily. At twenty dollars flat, you stop counting tokens, and counting tokens while you work is its own drain on concentration.
Two discounts get ignored constantly. The batch API takes fifty percent off both input and output tokens if you can wait for the results, which suits overnight jobs like crawling a client site or scoring a few thousand pages. Prompt caching charges cache hits at a tenth of the base input rate, so the second run against a large file costs far less than the first.
Then point the cheap model at the boring work. Opus 5 runs five dollars per million input tokens and twenty-five per million output, five times what Haiku 4.5 charges on both counts, and most of what you build will never need it. Parsing a file, renaming variables, and formatting a report all run fine on the smaller model. Save the expensive one for the logic you cannot reason through yourself.
