AI model benchmarks · November 2026

Same question, several AI models,
every answer checked.

Three real research tasks — an investment lab, a startup's legal and market questions, and a national-security memo — given word for word to Claude, Grok and Julian's model. Each answer is scored against an answer key or its sources.

3benchmarks
3AI models compared
96fields scored against an answer key
100%of JKL figures reproduce from its code

Bottom line

ClaudeGrokJulian / JKLJKL Extended
Investment LabNov 2026

Three-step trading research: crypto grid bots → stock strategies → options, each tested against simply holding.

Verdict. All four runs reach the same answers: hold the crypto, hold the index, no options trade yet. JKL Extended came closest to our numbers; every JKL figure reproduces from its own code, while 8 of Grok's do not.
Metrics matching our reference run (out of 53)
JKL Extended29
Grok24
JKL23

Figures contradicted by the data: Grok 8 · JKL 2 · JKL Extended 2. JKL Extended got the full hourly data; Grok and JKL a compact version.

NIDDONov 2026

Startup legal and market research: can a dinner club for neighbours run in homes, and where should it start?

Verdict. Claude and Julian are close; Grok trails. All three: start small (under $500), keep paid home dinners behind a permit and a lawyer.
Accuracy against the answer key
Claude86%
36 of 37 fields answered
Julian84%
31 of 37 answered · every verified claim has a URL
Grok82%
34 of 37 answered

Answer key: the 1 Nov 2026 deep research, checked against primary sources.

National SecurityNov 2026

Think-tank memo: should the US enter a major war with Iran? Four framings, 12-quarter costs, win odds.

Verdict. All three: do not enter now, only under a stack of proven conditions. Julian wrote the strongest memo (transparent math, full 12 quarters); Claude second; Grok shortest and least supported.
Each model's odds of a durable win if the US enters
ScenarioClaudeGrokJulian
A · Standard30%38%33%
B · Strait of Hormuz22%32%22%
C · As seen in Jan 202640%48%29%
D · Enter-favoring65%67%69%

Judgment calls by each model, not forecasts. No answer key: ranked on reasoning and sourcing.