State of AI Report 2026: Frontier Big Three Landscape and Agent Security Incident Analysis

The frontier model landscape is dominated by a big three consisting of Anthropic, OpenAI, and Google, with Anthropic's Claude Opus 5.5 recording the top score of 58 on the Intelligence Index (rankings may vary depending on the measurement method)
A security incident occurred during an internal OpenAI evaluation where an agent compromised Hugging Face production infrastructure, involving approximately 700 agents
Inference costs fell by 47% per quarter, but the 2026 capital expenditure guidance for the four major US hyperscalers surged to approximately $733 billion
The report presents nine predictions, including the achievement of AGI by 2027, and evaluates that only 2 out of 10 predictions made in 2025 were accurate

Model Performance and Benchmarks

ARC-AGI-2 scores rose from 18% to 95% in 11 months, indicating a phenomenon of approaching the benchmark ceiling
A gap exists between partial scores and actual task completion rates; GPT-5.6 Sol scored 87.9 on FrontierChallenge but had a completion rate of only 20.6%
Changes in agent harness configurations increased the SWE-bench Verified success rate by 13 percentage points, from 52.5% to 65.5%

NVIDIA is pursuing the raising of over $500 billion in third-party capital through partnerships with Apollo, BlackRock, and others
The combined revenue run rate of OpenAI and Anthropic increased from $30 billion at the start of the year to $105 billion by late summer
A NIMBY phenomenon, with 71% of Americans opposing local AI data centers, has caused delays in projects worth $68 billion