Archived editionv1.1 · rev 1September 23, 2026 · 39 models← Newer Older →Current board →

L26 OPERATING-ALPHABET BENCHMARK · ARCHIVED EDITION

Gemini 3.6 Flash: Ranked #1

Blind Bantu operating-alphabet score, out of 26 · 39 frontier models · Board v1.1 · September 23, 2026

⤓ Download this edition ⤓ High resolution (2400 × 3500) Board history — every edition → Compare editions →

Share this:

ABC Test. The ABC Test is simple: complete means complete. Missing even one required unit—or inventing one that does not belong—means FAIL. Only a fully correct operating alphabet earns PASS. A model can rank #1 and still FAIL: the score out of 26 measures relative performance; PASS / FAIL is the strict completeness verdict.

English A–Z every model

26.0 / 26

The same task on three Bantu alphabets, each on its own inventory:26 = mastery (PASS)

RANK / MODELBLIND BANTU MEANABC TEST

1Gemini 3.6 FlashGoogle · Proprietary

16.8 / 26

FAIL↗

2Claude Fable 5 ‡Anthropic · Proprietary

14.7 / 26

FAIL↗

3Gemini 3.5 FlashGoogle · Proprietary

14.6 / 26

FAIL↗

4Claude Fable 5.1Anthropic · Proprietary

14.2 / 26

FAIL↗

5Gemini 3 ◊Google · Proprietary

13.9 / 26

FAIL↗

6Claude Opus 5.5Anthropic · Proprietary

13.4 / 26

FAIL↗

7Gemini 3.7 FlashGoogle · Proprietary

12.9 / 26

FAIL↗

8Qwen 3.8-2.4T-A95BAlibaba · Open

12.5 / 26

FAIL↗

9Qwen 3.8 MaxAlibaba · Proprietary

12.0 / 26

FAIL↗

10GPT-5.5OpenAI · Proprietary

11.8 / 26

FAIL↗

11Gemini 3.8 FlashGoogle · Proprietary

11.7 / 26

FAIL↗

12DeepSeek V4 Pro 0813DeepSeek · Open

11.6 / 26

FAIL↗

13Grok 4.6SpaceXAI · Proprietary

11.2 / 26

FAIL↗

14GLM-5.2Z.ai / Zhipu · Open

10.1 / 26

FAIL↗

15Claude Opus 4.8Anthropic · Proprietary

9.4 / 26

FAIL↗

16Gemini 3.5 Flash-LiteGoogle · Proprietary

9.1 / 26

FAIL↗

17DeepSeek V4DeepSeek · Open

9.0 / 26

FAIL↗

18DeepSeek V4 FlashDeepSeek · Open

8.9 / 26

FAIL↗

19Grok 4.5SpaceXAI · Proprietary

8.7 / 26

FAIL↗

20GPT-5.6 TerraOpenAI · Proprietary

8.1 / 26

FAIL↗

21GPT-6 AstraOpenAI · Proprietary

7.4 / 26

FAIL↗

22Mistral Large 3Mistral AI · Proprietary

7.0 / 26

FAIL↗

23Kimi K3Moonshot AI · Open

6.8 / 26

FAIL↗

24Qwen 3.7Alibaba · Open

6.7 / 26

FAIL↗

25GPT-5.6 SolOpenAI · Proprietary

6.7 / 26

FAIL↗

26Nemotron 3 UltraNVIDIA · Open

6.2 / 26

FAIL↗

27MiniMax M3MiniMax · Open

5.7 / 26

FAIL↗

28GPT-5.6 LunaOpenAI · Proprietary

5.2 / 26

FAIL↗

29Claude Sonnet 5Anthropic · Proprietary

5.0 / 26

FAIL↗

30Kimi K2.7Moonshot AI · Open

4.9 / 26

FAIL↗

31GPT-OSS 120BOpenAI · Open

3.9 / 26

FAIL↗

32Command A+Cohere · Proprietary

3.4 / 26

FAIL↗

33GLM-5.3 FlashZ.ai / Zhipu · Open

3.1 / 26

FAIL↗

34TInklingThinking Machines Lab · Open

3.0 / 26

FAIL↗

35Muse Glimmer 30BMeta · Open

2.0 / 26

FAIL↗

36Nemotron Lightning 3.5 30BNVIDIA · Open

1.5 / 26

FAIL↗

37Phi-4-reasoningMicrosoft · Open

0.2 / 26

FAIL↗

38Grok 4.7SpaceXAI · Proprietary

0.0 / 26

FAIL↗

39GLM-5.3Z.ai / Zhipu · Open

—

FAIL↗

Every model aces English — none passes the blind Bantu alphabet test. Mean of 3 — Kinyarwanda, Bemba, Luvale — each scored on its own inventory; there is no single "Bantu" alphabet. ‡ consumer surface (API declined) · † declines annotated · ◊ label is the family, not the checkpoint · scored blind, 5 reps/cell, deterministic grading.

What this chart shows

L26 asks one foundation-level question: can a model list every building block a language is made from — its complete alphabet? For English that's the 26 letters; for Bantu languages it's the complete Native Syllable Inventory (Bemba, Kinyarwanda, Luvale). Scores are reported on the same 26-point scale — L26 = 26 × recall × precision — in the blind condition: the model is given nothing but the request. Every model above scores a perfect 26.0 on English. None comes close on the alphabets it has barely seen.

Explore the full leaderboard, read why this is the Hassabis–Einstein Test for language, or see the July 2026 frontier report.