OpenAI wants to measure AI by results, not tokens. Africa's bill reads differently
OpenAI's new "useful intelligence per dollar" scorecard asks whether AI pays for itself in completed work. Applied to a market where retries cost real money and $42 is a month's minimum wage, the metric exposes a gap the framework itself doesn't price in.
Forty-two dollars. That is roughly what Nigeria's ₦70,000 monthly minimum wage converts to at current exchange rates — barely enough, employment researchers at Employsome and RemotePeople both note, to cover a single week of basic costs in Lagos, let alone a month. It is also close to what a single developer could spend in an afternoon running a moderately busy AI agent on OpenAI's cheapest new model, if a few of its calls don't land clean the first time.
That collision is the real story behind a document OpenAI published on July 17: "A scorecard for the AI age," written by chief financial officer Sarah Friar. Its argument is simple and, on its face, sensible. Companies have been measuring AI adoption by the wrong yardstick — seats bought, subscriptions signed, cost per token — when the only number that matters is cost per successful outcome: how much it actually costs, all in, to get one unit of real work done correctly. Friar calls it "useful intelligence per dollar," built on four questions: how much useful work is getting done, what a successful task actually costs, how often the AI gets it right without a human stepping in, and whether each dollar of compute buys more work as usage scales.
The timing is not incidental. A week earlier, on July 9, OpenAI had taken its newest model family, GPT-5.6, out of a government-reviewed limited preview and into general release. Rather than one model, it shipped three: Sol, the flagship, at $5 per million input tokens and $30 per million output; Terra, a mid-tier workhorse at $2.50 and $15; and Luna, the cheapest and fastest, at $1 and $6 — figures confirmed on OpenAI's own developer pricing page and cross-checked against three independent pricing trackers (Codersera, eesel, and BenchLM), all reporting identical numbers. OpenAI's pitch, echoed by CEO Sam Altman in a CNBC interview the same week, is that Sol reaches new coding benchmarks using 54% fewer output tokens than a rival flagship model — meaning the same task should, in theory, cost less to finish even where the sticker price hasn't dropped.
Why cost-per-token was always the wrong question for African builders
Himilo has covered this terrain before: African developers building on frontier US models are billed in dollars, for infrastructure priced for enterprises with stable power and broadband. Friar's scorecard, without meaning to, sharpens exactly where that gap bites. Her framework treats "needs correction" and "needs escalation" — a task that requires a retry, or a human to finish it — as a dependability problem to be optimized away over time. In much of Nigeria, Kenya, or South Africa, though, a dropped mobile connection or a brownout mid-request isn't a dependability edge case measured in single-digit percentages; it's a routine tax on every API call, and every retry is billed again at full price. A workflow that would clear OpenAI's own "successful task" bar in San Francisco can quietly cost two or three times as much in Lagos, for reasons that have nothing to do with the model's intelligence.
That distinction matters because it is precisely the kind of context an outsider wouldn't have. A $1-per-million-token Luna call looks trivial against an American mid-market salary. Measured against Nigeria's $42 monthly minimum wage — or Kenya's roughly $2,363 GDP per capita for 2025, per World Bank data — the same call sits in a completely different part of a household or small-business budget. A founder billing a local customer in naira, but paying OpenAI in dollars pegged to US enterprise economics, absorbs a currency and connectivity risk that never appears in Friar's four questions, because the scorecard was written for a market where the dollar is the local currency and the grid rarely fails.
The moat that isn't priced in
None of this makes GPT-5.6 a bad deal for African startups — Terra's roughly 2x cost cut over the previous flagship, if it holds up in production, genuinely lowers the floor for building here. But it does mean OpenAI's new "useful intelligence per dollar" logic, exported wholesale, undercounts the real cost of doing useful AI work in markets where the model is rarely the unreliable part of the stack — the connection and the currency are. Any African founder adopting the framework to pitch investors should build in a wider margin for retries and FX volatility than the scorecard assumes, and treat Friar's own four questions as a checklist to interrogate, not a formula to import unchanged.
OpenAI has not, as of this writing, published region-specific reliability or cost data for Africa — a gap worth watching as the company expands its enterprise sales push on the continent. Until it does, the most useful metric for a Lagos or Nairobi engineering team may not be dollars per completed task, but dollars per completed task after the network drops the connection once.
