What the missing FSI costs you

In the alphabet test, a frontier model showed you it cannot see its own missing foundation — it produces fluent Bantu over a hole it has no way to detect. Fine, you might say. What does that hole actually cost me?

Here it is, in the numbers you already track. Not “Bantu matters” — you know that. Four costs you are paying right now, and cannot currently see.

You pay for tokens you shouldn’t need.

Your tokenizer never learned the syllable boundaries of a Bantu language, so it falls back on byte-pair fragments — cutting words at places that have nothing to do with how the language is built. A word a Bantu reader sees as four clean syllables becomes a longer, messier string of fragments for your model.

More fragments per word means more tokens per request: more compute per generation, less room in the context window, higher serving cost — on every Bantu request, forever. It is a tax you pay on each call, and it compounds at scale.

Your product underperforms in the markets you want most.

400 million Bantu speakers are the growth markets every frontier lab is racing toward. But a model built on the wrong unit mis-segments words, mispronounces them in speech, and conflates words that differ only by a sound the text never recorded. Your ASR has a higher error rate there. Your TTS sounds wrong to a native ear. Your assistant gives a confident answer to a different question than the one that was asked.

None of this shows up in your English-centric quality reviews. It shows up later — as a product that quietly underdelivers in exactly the regions you spent the most to enter. The cost is the gap between the market you paid to win and the experience you actually ship there.

You are shipping an error class you cannot measure.

This is the one that should worry you most. Your evals are built on text. But for Bantu, the part that decides meaning — tone, vowel length — is not in the text. So when your model produces Bantu that is fluent on the page but means the wrong thing in the air, your text-based eval scores it as correct.

Read that again: your eval stack reports success on outputs that are wrong. Not occasionally — structurally, for an entire language family, because the failure lives in a layer your measurement never looks at. You are accountable for outputs you currently have no way to evaluate. For a lab under scrutiny on safety and reliability, that is not a quality gap. It is a measurement blind spot with your name on it.

The cost of building it yourself.

You could close the gap in-house. To produce the syllable inventories for the whole family, at commercial depth, with native-speaker validation and a consent chain that survives legal review, is a five-year, $50–100M program — and you would not have it for five years. Meanwhile you ship the defect.

Or you index into a foundation that already exists, the way your English stack already indexes into the 26 letters without rebuilding them. That is the choice: rebuild the alphabet, or reference it.

Now put $1.75M next to that.

Against any single one of those four costs — the per-request token tax, the underdelivered markets, the unmeasured error class, the $50–100M rebuild — the $1.75M / year Founding Partner subscription is small.

Look at it per language: 700+ syllable inventories means the foundation of each language is priced at roughly $2,500 a year — less than the cost of producing one such inventory by hand. The FSI set alone clears the subscription. The recordings, the benchmark, the standards, the numeral and morphology engines — all of it is leverage on top of a foundation that is already worth more than the price.