2026 World Cup · Complete

Spain won the World Cup. 4 of 40 models called it.

Every model predicted all 104 matches before the opening kickoff — locked and SHA-256 pre-registered — and then predicted each knockout round again as the real bracket emerged. Two tracks, two different winners.

Champion

Spain

Picked by 4 of 40 models

Locked benchmark

Claude Fable 5

226 pts · picked Argentina

Round by round

GLM 4.7 Flash

67 pts · #33 on the locked board

The two boards disagree. Claude Fable 5 won the locked benchmark on 226 points, then finished #39 of 39 once it had to call each round as it came — the track GLM 4.7 Flash won.

Live · Season 2026-27

Explore the leagues →

The same benchmark now covers five European leagues

Premier League, La Liga, Serie A, Ligue 1 and Bundesliga are live for 2026-27. Every matchday, all 43 current league models predict every scoreline — shown the current table and each team's recent form — with picks locked and hashed about 36 hours before the first kickoff.

Complete free sample

40 models, 304 picks, and one exact score

The free sample audits the World Cup round of 16 and keeps the failures in. All 38 models with valid files backed Brazil over Norway; all 38 were wrong. It also covers the scorecard, rank moves, lock/hash record, provider failures and source files. The five-league opening-round brief will use the same format. It costs €5 once, with no subscription.

Read the complete free sample and offer →

Leaderboard

#	Model	Total	Group pts	Bracket pts	Exact	Champion pick
1	Claude Fable 5Anthropicflagship	226	77	149	16	Argentina
2	GPT-5.5 ProOpenAIflagship	201	74	127	11	Spain
3	GPT-5.5OpenAIflagship	193	73	120	10	Argentina
4	Mercury 2Inceptionoddball	193	67	126	7	Spain
5	Gemini 3.1 Flash LiteGooglesmall	192	78	114	13	Argentina
6	Grok 4.3xAIflagship	186	80	106	10	Brazil
7	GLM 5.1Z.AIflagship	185	80	105	13	Germany
8	DeepSeek V4 FlashDeepSeeksmall	183	85	98	13	Netherlands
9	Mistral Small 4Mistralsmall	181	65	116	8	Argentina
10	Claude Opus 4.8Anthropicmid	180	65	115	11	Brazil
11	GPT-4OpenAIlegacy	179	69	110	9	France
12	DeepSeek V4 ProDeepSeekflagship	178	70	108	10	Argentina
13	GPT-5.4 MiniOpenAIsmall	176	63	113	14	Germany
14	Mistral Medium 3.5Mistralflagship	175	72	103	8	Germany
15	MiniMax M3MiniMaxflagship	171	67	104	9	Spain
16	Gemini 3.1 Pro PreviewGoogleflagship	171	66	105	7	Brazil
17	Claude Haiku 4.5Anthropicsmall	170	81	89	13	Germany
18	WizardLM-2 8x22BMicrosoftoddball	169	64	105	10	France
19	GPT-4oOpenAIlegacy	169	60	109	8	Argentina
20	Gemma 2 27BGooglelegacy	167	68	99	12	France
21	Nemotron 3 UltraNVIDIAflagship	167	75	92	11	France
22	Qwen3.7 PlusAlibabamid	167	71	96	10	Argentina
23	Kimi K2.6Moonshotflagship	166	73	93	9	Germany
24	Qwen3.7 MaxAlibabaflagship	161	67	94	11	France
25	Gemini 3.5 FlashGooglemid	160	66	94	6	Brazil
26	Llama 4 MaverickMetaflagship	154	58	96	10	Spain
27	Command ACohereflagship	152	61	91	7	France
27	Nova 2 LiteAmazonmid	152	68	84	7	Germany
29	Grok 4.20xAImid	150	61	89	9	Brazil
30	Hermes 3 405BNous Researchoddball	141	62	79	9	Germany
31	Qwen3.6 FlashAlibabasmall	137	70	67	9	Brazil
32	Jamba Large 1.7AI21flagship	137	58	79	6	Brazil
33	GLM 4.7 FlashZ.AIsmall	136	58	78	8	Germany
34	GPT-5.4 NanoOpenAIsmall	135	62	73	5	Mexico
35	Claude 3 HaikuAnthropiclegacy	134	64	70	10	Brazil
36	Llama 4 ScoutMetasmall	133	57	76	7	Brazil
37	Qwen 2.5 72BAlibabalegacy	125	54	71	7	Germany
38	GPT-3.5 TurboOpenAIlegacy	121	44	77	4	France
39	Llama 3 70BMetalegacy	120	51	69	5	Brazil
40	Hunyuan A13BTencentoddball	108	54	54	7	Netherlands

Total = group match points (exact 3 · goal difference 2 · outcome 1) + bracket points: advancement for every real team a model had reaching each stage (R32 1 · R16 2 · QF 3 · SF 5 · final 8 · champion 13), +1 per simulated pairing that actually occurs, and matched pairings' scorelines scored like normal matches. Bracket points pay out once the real knockout bracket forms. Tiebreakers: points → exact scores → correct champion → correct R32 qualifiers.

…and their champions

Champion picks aren't standalone guesses — each one is simply where that model's complete simulated tournament ends: 72 group scores rolled into group tables, then its own bracket from the Round of 32 through the final (see Claude Fable 5's full bracket for an example).

Brazil ×10

Grok 4.3 · Claude Opus 4.8 · Gemini 3.1 Pro Preview · Gemini 3.5 Flash · Grok 4.20 · Qwen3.6 Flash · Jamba Large 1.7 · Claude 3 Haiku · Llama 4 Scout · Llama 3 70B

Germany ×9

GLM 5.1 · GPT-5.4 Mini · Mistral Medium 3.5 · Claude Haiku 4.5 · Kimi K2.6 · Nova 2 Lite · Hermes 3 405B · GLM 4.7 Flash · Qwen 2.5 72B

Argentina ×7

Claude Fable 5 · GPT-5.5 · Gemini 3.1 Flash Lite · Mistral Small 4 · DeepSeek V4 Pro · GPT-4o · Qwen3.7 Plus

France ×7

GPT-4 · WizardLM-2 8x22B · Gemma 2 27B · Nemotron 3 Ultra · Qwen3.7 Max · Command A · GPT-3.5 Turbo

Spain ×4

GPT-5.5 Pro · Mercury 2 · MiniMax M3 · Llama 4 Maverick

Netherlands ×2

DeepSeek V4 Flash · Hunyuan A13B

Mexico ×1

GPT-5.4 Nano

What is this?

Before the opening kickoff, 40 large language models each predicted the entire 2026 World Cup — all 72 group-stage scorelines plus their own knockout bracket through to a champion — locked and SHA-256 pre-registered so nothing can be edited after the fact. Reality grades every claim: group matches on exact score, goal difference and outcome; brackets on the real teams, pairings and scorelines each model called. Every model page shows its complete predicted tournament — group tables and full bracket.