AGI Benchmarks & Frontier Models Intelligence Report (September 2026)

Humanity's Last Exam (HLE) PhD Level

Claude Fable 5.1

65.0% Accuracy 2,500 questions

Archetype No Cap King

ARC-AGI-3 (Dynamic Agent) Game Solving

GPT-6 Astra

99.95% Solve Rate Historic Leap

Archetype ARC Final Boss

SWE-bench Pro (Private Repos) Unseen Code

Claude Fable 5.1

58.4% Pro (95.2% Verified)

Contamination Check Zero Leakage Dev

Throughput & Scale Fast & Cheap

Gemini 3.8 Flash

185 tok/s 2M ctx @ $0.61/M

Archetype The Speed Demon

Intelligent Decision Engine • Fall 2026

Choose the Best Model for Your Specific Task

There is no longer a single general winner in September 2026. Different models have diverged sharply into distinct architectural specialties: agentic exploration, multi-file software engineering, zero-latency throughput, or PhD scientific rigor. Select your primary workload below to view the tailored recommendation.

Frontier AGI Benchmarks Master Table

Click any column header to sort. Filter models or export the dataset to CSV.

Quick Filters:

Model	AGI Index	HLE (%)	ARC-2 (%)	ARC-3 (%)	SWE Pro (%)	GPQA (%)	FrontierMath	LCBench (%)	Arena Elo	Context	Price/M	Speed

Benchmark Saturation & Anti-Contamination Analysis (Sep 2026)

⚠️ Saturated Benchmarks

MMLU, GSM8K, ARC-1, SWE-bench Verified: Frontier models score 95-100%. These benchmarks no longer differentiate top models and suffer from heavy training set memorization. Always test against second/third generation suites.

🛡️ Active Unsolved Frontiers

Humanity's Last Exam (HLE) & FrontierMath: With top models scoring 58-65% on HLE and 31-42% on FrontierMath, these tests remain the most reliable indicators of expert scientific and mathematical cognition.

🎮 The Interactive Frontier

ARC-AGI-3 & GAIA Level 3: Moving away from static prompts toward dynamic game loops and multi-step browser tool invocation. This evaluates autonomous learning and trial-and-error hypothesis testing.

AGI Benchmarks & Model Intelligence

Choose the Best Model for Your Specific Task

Comparative Benchmark Explorer

ARC-AGI Series Comparison

ARC-AGI Fluid Intelligence Progression

SWE-bench Verified vs Pro Discrepancy

8-Axis Capability Spider

Toggle Active Models

Price vs. AGI Intelligence Pareto Frontier

Frontier AGI Benchmarks Master Table

Benchmark Saturation & Anti-Contamination Analysis (Sep 2026)

Frontier Model Personas: No Cap, Pure Signal