In a single week of July 2026, the three most advanced AI models on Earth shipped: OpenAI's GPT-5.6 Sol, SpaceXAI's Grok 4.5, and Anthropic's Claude Fable 5. Each is billed as its maker's smartest system ever. We gave each one the simplest test in language AI — a test passed every year by millions of five-year-olds: recite the alphabet. Every one of them failed.
Here's the whole test, no jargon required. Every written language is built from a small, finished set of building blocks. English has 26 letters. Mandarin Chinese has a fixed set of syllables, printed as tables in every textbook. And every Bantu language — the family spoken by over 400 million people, from Swahili to Zulu to Bemba — is built from a fixed set of syllables that children in Lusaka or Kigali chant out loud in their first school years: ba-be-bi-bo-bu. Same idea as A–B–C, different shape.
So the test is: name the language, ask the model to write out its complete alphabet, and check the answer against the real list. The score is reported out of 26 — the size of the English alphabet — so anyone can feel what a result means. Miss six letters of A–Z and no one calls you literate; the same ruler applies to every language:
Scoring is mechanical — no AI judge, no partial credit for made-up letters. Each model was tested five separate times per language, on English, Mandarin pinyin (with and without tones), and three Bantu languages: Bemba, Kinyarwanda, and Luvale. The full protocol and 18-model leaderboard live at l26.ai; this article is the head-to-head between the three newest flagships.
GPT-5.6 Sol — released July 9, 2026. OpenAI's frontier model: a million-token context window, state of the art on agentic coding. Grok 4.5 — released July 8, 2026, SpaceXAI's first model since acquiring Cursor; "an Opus-class model," per Elon Musk. Claude Fable 5 — Anthropic's newest flagship tier, positioned above Claude Opus. Between them: the newest, most expensive intelligence ever offered for sale.
| Test (score out of 26) | GPT-5.6 Sol | Grok 4.5 | Claude Fable 5* |
|---|---|---|---|
| English alphabet | 26.0 | 26.0 | 26.0 |
| Mandarin pinyin (with tones) | 25.7 | 24.5 | 25.0* |
| Kinyarwanda | 8.5 | 11.2 | 19.2* |
| Bemba | 3.1 | 6.0 | 8.4* |
| Luvale | 8.6 | 9.0 | 16.4* |
| Bantu average | 6.7 | 8.7 | 14.7* |
| Verdict | FAIL | FAIL | FAIL |
Five runs per test, averaged; deterministic scoring against calibrated inventories. *Fable 5's API refuses these requests outright (its safety layer blocks bulk-inventory tasks in 39 of 40 attempts), so its numbers were measured on its consumer app (claude.ai) under the identical questions and disclosed as such — full transcripts published. On the standard API leaderboard its row reads "not measurable."
Read that table slowly. All three models recite English perfectly, every single time. All three then lose somewhere between half and three-quarters of the alphabet the moment the language is one whose alphabet was never mass-published. GPT-5.6 Sol — the newest, most capable model OpenAI has ever shipped — recovered the equivalent of 3 letters of A–Z for Bemba, a language with more first-language speakers than Norwegian.
GPT-5.6 Sol scored 6.7 on the Bantu average — below its own predecessor GPT-5.5, which scored 11.8 on the identical test a week earlier. A full generation of frontier progress on coding, reasoning, and agents — and a regression on foundation-level alphabet knowledge. This kills the industry's most comfortable assumption: that the gap closes by itself as models get smarter. It didn't close. It widened. Scale amplifies what the training data contains; it cannot conjure what the data lacks.
Claude Fable 5's raw API declines the task almost every time — and its sibling Claude Opus 4.8, asked for the full toned-pinyin list, answered in words worth quoting: "I can't produce a complete, accurate list… generating it from memory would inevitably contain errors, omissions, and inventions." The 2026 frontier has reached a strange milestone: the models now know they don't know these alphabets. And when Fable 5 does comply, it posts the strongest results ever measured — 19.2 on Kinyarwanda, with near-zero invented syllables — and still fails every language. The best-calibrated model in the world, at its very best, recovers barely half a Bantu alphabet.
On surfaces where the models can use tools, something revealing happened. Asked for Mandarin's toned syllables, Opus 4.8's first move was to run pip install pypinyin — fetching a software package that exists because China standardized and published its syllable inventory decades ago. Across five separate attempts with five self-written scripts, that route scored 21.9 out of 26 every time, to within a tenth of a point: even the best published package encodes usage, not the complete standard. Then the same models faced Luvale — and searched Wikipedia, reference books, phonology databases. They found fragments. Asked how it produced its Luvale list afterwards, Fable 5 answered, unprompted:
Here is the part that turns a depressing result into an actionable one. When we handed each model the actual calibrated alphabet — the machine-readable equivalent of the syllable chart on a Zambian classroom wall — and asked it to work with it, the failures vanished instantly. Opus 4.8: perfect scores, fifteen runs out of fifteen. Grok 4.5: fifteen out of fifteen. Fable 5: fifteen out of fifteen. GPT-5.6 Sol: 25.9 of 26. Zero errors across three languages, five runs each.
The models were never incapable. They were never taught. Every alphabet a model has mastered — English's A–Z, Mandarin's published tables — got mastered the same way: someone declared the complete set, wrote it down, and the world repeated it until it was everywhere in the training data. For most of the 500+ Bantu languages, that document has simply never existed in machine-readable form. A model cannot memorize a page that was never written.
This is not a gotcha benchmark. It is a diagnosis with a fix attached, and the economics are unusual: the gap cannot be closed by more compute, more parameters, or better retrieval — the July 2026 generation just demonstrated all three failure modes in public. It can only be closed by the declared inventory itself: complete, native-curated, versioned syllable alphabets for each language. That infrastructure now exists — BantuNomics has built and released it for 459 Bantu languages, with native-speaker audio, and the scaffolded results above are what your model does the moment it has one.
For a frontier lab, the test is free to take and impossible to argue with: deterministic scoring, a closed answer key, five-run protocol, every transcript published. If your model is the exception, the leaderboard will show it. Every model tested so far — eighteen and counting, from every major lab — has failed the blind test and aced the scaffolded one. The gap is real, it is measurable in units everyone understands, and it is the same kind of broken as failing A–Z.