Research & findings.
Reports, cohort studies and essays based on published alphabet evaluations.
Kimi K3, Inkling, GPT-5.6 Sol, Claude Fable 5, Grok 4.5: Five July Flagships, One Alphabet Test — Open and Closed AI Both Failed
Ten days in July 2026 produced five flagship AI releases — the largest open model ever (Kimi K3), the leading U.S. open model (Inkling), and three closed frontiers (GPT-5.6 Sol, Claude Fable 5, Grok 4.5). We scored them all on the same alphabet test. Open or closed, every one failed — and the open-vs-closed gap turns out to be a rounding error next to the real divide.
Read publication ↗GPT-5.6, Fable 5, Grok 4.5 Compared on Alphabet Tests — and the Results Are In: All Three Failed
In July 2026 we gave the three newest frontier AI models the simplest test in language AI: recite a language's complete operating alphabet.
Read publication ↗The newest frontier models just took the alphabet test. The gap didn't close.
GPT-5.6 Sol and Grok 4.5 shipped in July 2026. Within days we scored them on L26, the operating-alphabet benchmark. The newest generation did not close the foundation gap — one flagship scored below its predecessor — and the data says the gap will not close on its own.
Read publication ↗Google Shipped Two Newer Flash Models. Both Got Worse at the Alphabet Test.
Gemini 3.6 Flash leads the L26 board. Google then shipped 3.7 Flash and 3.8 Flash — and both scored LOWER than the model they replaced. On Bemba the newest Flash recovers a third of what its own predecessor recovered, while still scoring a perfect 26/26 on English. Deliberation time halved across the same releases.
Read publication ↗Nine New Frontier Models Took the Alphabet Test. The Best One Came Ninth.
Nine models from five labs — OpenAI, DeepSeek, Z.ai, Alibaba, Meta and NVIDIA — arrived in the first days of September 2026. None displaced a single model from the top five, a July release still leads, and at two labs independently the newer model scores worse than the one it replaced. Every one of them aces English.
Read publication ↗They Are Calling It AGI. It Could Not Recite an Alphabet.
OpenAI's GPT-6 Astra, released September 3, 2026 and widely described as possible AGI, scores a perfect 26/26 on the English alphabet. Asked for the syllables of Luvale, it declined five times out of five. Handed the list, it rebuilt Bemba, Kinyarwanda and Luvale perfectly. A language-equity result on the most basic test there is.
Read publication ↗