The Discount in a Currency Your Language Can't Spend

The number everyone repeats is the price of an inference: roughly $0.50 down to $0.03 a query, a fall of more than 90% in eighteen months. Real number. Genuinely good news. Also, quietly, the wrong unit to celebrate. That price is per token — and a token is not a unit of meaning. It is an artifact of the text a tokenizer was trained on, which was overwhelmingly English. The discount is denominated in a currency rigged against the languages most of the unconnected world actually speaks.

Start with what a tokenizer does to a sentence in Shona. Byte-level and subword tokenizers build their vocabulary by counting frequent substrings in a training corpus. When that corpus is mostly English, the model gets a dense, efficient vocabulary for English and a coarse, brittle one for everything else. Feed it an agglutinative, under-represented African language and the same idea shatters into far more pieces. The research calls this fertility — tokens per word — and a recent audit across 16 African languages, 9,000 questions, and 10 models found that fertility "reliably predicts accuracy" 1. The more your language fragments, the less the model understands it. Spoken weight is no shield: isiZulu, spoken by over 12 million South Africans, and Hausa, spoken by more than 70 million across West Africa, are both classed "low-resource" in AI terms, because they leave a small digital footprint to scrape 11.

The penalty is easiest to read in money. One analysis of commercial APIs found Greek encoding the same content in 42 tokens — 6× the English count — turning a chatbot that costs about $16,425 in English into roughly $98,550, an $82,000 annual gap for identical functionality 9. Greek is a relatively well-resourced language. The pattern it exposes — Latin-script English cheapest, everything else taxed — lands hardest on languages with even thinner training data.

The tax has more than one face. More tokens per meaning is not just a billing line; it compounds in four directions at once. Cost — you pay the multiplier on every call. Latency — more tokens to generate, over a network already degraded. Effective context — the same window holds less of the actual conversation, because so much of it is spent on fragments. And quality — the fertility-accuracy link in 1 means the fragmented language is also the one the model understands least. One mechanism, four deficits, all landing on the same users — the ones for whom mobile data already costs 10–20% of monthly income.

Here is the hinge, the part the headline cost-collapse hides. The tokenizer sits upstream of the inference price. So when cost-per-token falls, it falls for everyone — but the ratio between languages does not move. Discount and penalty travel together. English queries drift toward free; the multiplier on Shona rides the discount down at exactly the same rate. The absolute gap narrows, which is what gets quoted. The relative tax is untouched — and at sub-cent margins, relative is what bites. This is the interface deficit showing up one layer below the interface, in the encoding itself. Structural, not incidental. The providers optimizing for aggregate token volume have no reason to fix it — and the drift can run the wrong way. Claude Opus 4.7, launched April 2026, kept the same rate card but shipped a tokenizer that can produce more tokens for the same input, so effective cost per request can rise even when the price didn't change 14. The rate card is not the bill.

The compute math makes the stakes worse than linear. Because transformer cost scales with the square of sequence length, a doubling in tokens does not double cost — it roughly quadruples training cost and time 1. The tax is convex. The languages already paying most pay disproportionately more as you scale.

So what closes the gap? Not the next price cut — that just walks both numbers down in lockstep. The real levers are upstream, where the damage is done: tokenizers extended or rebuilt for the target language, vocabulary adaptation and continued pretraining, and tokenizer-transfer methods that lift low-resource performance without retraining from scratch 4. Purpose-built African tokenizers already beat global multilingual ones — by an average of 1.70% on sentiment classification in one study 8. Small, but it points the right way. And the most practical lever for anyone building today: choose models by tokens-per-meaning in your language, not by headline price.

That is the systems point. For these markets the cost frontier was never the model — it is the encoding and the last mile. The cheapest token is not the one you negotiated down. It is the one you never had to spend, because the language was encoded fairly in the first place.