Research
Which AI Has the Highest IQ? How Machine Intelligence Is Measured in 2026
Benchmark figures are third-party or vendor-reported. Not tested by this site.
AI models don't have biological IQs, so how do we rank the smartest AI? Compare November 2026 scores across Humanity's Last Exam, GPQA, ARC-AGI-3, and Epoch AI's Capabilities Index.
At a glance
- Type
- Research
- Topics
- AI IQ, AI benchmarks, ARC-AGI-3
- Evidence
- Third-party research
- Written by
- Liyam Flexer
AI models do not have an official human IQ score, because traditional IQ tests measure human psychometric traits and are easily memorized during model pre-training. Instead, researchers measure "machine IQ" using contamination-resistant reasoning and fluid-intelligence benchmarks. As of November 2026, three models trade the title of smartest AI depending on the test:
Quick Answer (TL;DR)
Why Standard Human IQ Tests Do Not Work on AI
When viral charts claim a chatbot scored "140 or 155 on an IQ test," they are usually administering standardized human pattern tests (such as Raven’s Progressive Matrices or Mensa practice exams). Those numbers are misleading for two reasons:
- Training Data Contamination: Millions of logic puzzles, analogies, and psychometric test explanations already exist in public web scrapes used to train large language models.
- Spiky Cognitive Profiles: An AI model can solve PhD-level quantum mechanics problems in seconds while failing a simple novel grid puzzle that a 10-year-old solves intuitively.
To measure true reasoning power—what psychometricians call fluid intelligence—AI researchers rely on adversarial, expert-authored benchmarks.
The 2026 "Machine IQ" Leaderboard: Top Reasoning Benchmarks
| Benchmark | What Cognitive Skill It Measures | Why It Resists Memorization | Current Leader (Nov 2026) |
|---|---|---|---|
| ARC-AGI-3 | Fluid Intelligence & Novel Adaptation: Solving unseen interactive environments on the fly | Tasks require zero-shot abstraction in dynamic environments never seen in training | OpenAI GPT-6 Astra (state-of-the-art reported) |
| Epoch AI Capabilities Index | Composite Frontier Intelligence: Weighted synthesis of top reasoning, math, and coding evals | Continuously updated across held-out, high-difficulty evaluations | 1. GPT-6 Astra (169); 2. Claude Opus 5.5 (168) |
| Artificial Analysis Index | Applied Reasoning & Knowledge: Standardized multi-domain intelligence and coding suite | Evaluated independently under uniform API parameters | 1. Gemini 4 Carbon (59); 2. Claude Opus 5.5 (58); 3. GPT-6 Astra (57) |
| FrontierMath Tier 4 | Original Mathematical Discovery: Research-level problems crafted by professional mathematicians | Newly authored proofs requiring hours or days for human specialists | OpenAI GPT-6 Astra |
| Humanity's Last Exam & GPQA | Expert Academic Reasoning: 2,500+ ultra-hard questions across 100+ graduate disciplines | PhD-level questions screened to defeat web search and shallow recall | Claude Fable 5.1 / GPT-6 Astra / Gemini 4 Carbon |
| OSWorld 2.1 & Terminal-Bench 4.0 | Applied Executive Function: Multi-step planning and error recovery in live software | Graded by actual system state changes rather than text answers | Claude Opus 5.5 (81.8% OSWorld / 66.4% Terminal-Bench) |
What Drives High "AI IQ" in 2026?
The leap in reasoning scores across 2025–2026 comes from test-time compute scaling (also called "thinking effort"). Rather than emitting the first plausible word immediately, frontier models like Claude Opus 5.5, GPT-6 Astra, and Gemini 4 Carbon generate internal chains of thought—testing hypotheses, catching logical contradictions, running code interpreters to verify calculations, and backtracking before delivering a final answer.
Keep reading
Evidence in this article Third-party research Work published by researchers or institutions other than us or the vendor. We link it; we have not reproduced it.
Frequently asked questions
What is the highest IQ ever recorded for an AI?
AI systems do not receive valid human IQ scores. While frontier models routinely score above 130–145 on unverified online IQ quizzes, researchers evaluate true machine intelligence using ARC-AGI-3, FrontierMath, and Epoch AI’s Capabilities Index, where GPT-6 Astra (169) and Claude Opus 5.5 (168) lead in November 2026.
Which AI is smartest for math and science?
OpenAI’s GPT-6 Astra and Anthropic’s Claude Fable 5.1 lead graduate-level mathematics and scientific reasoning benchmarks, with Astra reporting state-of-the-art results on FrontierMath Tier 4.
Does higher reasoning ability mean fewer hallucinations?
Higher reasoning effort significantly reduces logical and mathematical errors because the model checks its own intermediate steps, though no model is completely immune to hallucinations without external grounding and verification.
Why don't human IQ tests work on AI?
Millions of logic puzzles, analogies and psychometric test explanations already sit in the public web scrapes used to train language models, so scores are contaminated, and AI models have spiky cognitive profiles that a single IQ number hides.
What drives high AI IQ in 2026?
Test-time compute scaling, also called thinking effort: frontier models generate internal chains of thought, test hypotheses, catch logical contradictions, run code interpreters to verify calculations and backtrack before answering.
Which model leads which benchmark in November 2026?
GPT-6 Astra leads Epoch AI's Capabilities Index (169), ARC-AGI-3 and FrontierMath Tier 4 as reported; Gemini 4 Carbon leads the Artificial Analysis Intelligence Index (59) and arena.ai (1528); Claude Opus 5.5 leads OSWorld 2.1 (81.8%) and Terminal-Bench 4.0 (66.4%).
Related reading
- Category: Research & knowledge
Sources
- arena.ai text leaderboard arena.ai.
- Artificial Analysis model leaderboard Artificial Analysis.
- Epoch Capabilities Index Epoch AI.
- ARC Prize ARC Prize Foundation.
- Humanity's Last Exam Center for AI Safety and Scale AI.
- Terminal-Bench Terminal-Bench.