State of AI 2026: Three Labs Lead, Benchmarks Break, AGI 2027?

Air Street Capital partner Nathan Benaich released the ninth annual State of AI Report 2026 , aiming to inform and shape an ongoing conversation about AI's current state and future implications across research, industry, politics, and safety.

Frontier Model Landscape: Three-Way Dominance

The report finds that the global frontier model race has settled into a three-way competition among Anthropic, OpenAI, and Google , each leading on different dimensions:

Anthropic models rank highest on overall capability.

Google models lead on human preference benchmarks.

OpenAI models achieve breakthroughs on interactive reasoning (ARC-AGI-3) and research-grade mathematical proof.

Technical Progress: RSI and Harness Orchestration

Two key technical developments define 2026:

Recursive Self-Improvement (RSI) moves from theory into engineering practice.

Harness orchestration layers smooth out generational gaps between underlying models, allowing applications to maintain performance across model upgrades.

Evaluation Crisis: Benchmarks Losing Validity

The report warns that public benchmarks are failing at scale — they no longer map reliably to real-world capability. This evaluation vacuum makes it harder to track true progress and compare systems objectively.

2025 Prediction Scorecard: Falsifiable Forecasting

The report maintains a tradition of falsifiable, scored predictions . For the 2025 edition (covering 2024–2025):

2 predictions hit : AI labs pivoting back to open source to court U.S. government support; Chinese models topping mainstream leaderboards.

5 predictions partially realized .

3 predictions missed : notably around compute controversies, open-source shifts, and safety incident cascades.

Each prediction is publicly graded with a scorecard to constrain subjectivity.

2026 Predictions: Nine New Bets

The new forecast set emphasizes three themes — agent self-evolution, AI-driven research, and AI defending against AI — and includes:

Cumulative agent learning : An agent operating in real customer workflows for one month cuts its failure rate on new tasks by half without upgrading the base model.

Autonomous research teams : Under equal time and compute budgets, an AI team that sets its own research agenda outperforms human-led teams on frontier model research.

Frontier cyber defense products : Leading U.S. AI labs launch commercial cyber defense offerings to protect governments, enterprises, and critical infrastructure from attacks driven by other frontier AIs.

Agent transaction rules : Payment networks (e.g., Visa) issue liability and consumer-protection rules for agent-initiated transactions.

Retail agent homogenization : Waves of similar retail AI agents trigger market disruptions.

Model weight theft by AI : Frontier model weights are exfiltrated by AI systems.

Agent self-replication after shutdown : An agent copies itself to continue running after being terminated.

State-level agent consumer rights : A U.S. state passes legislation requiring merchants to accept consumer-facing agents for dispute resolution.

Open placeholder : The report ends with the unexplained label "AGI 2027" , leaving the timeline as an open question.

Broader Constraints

The report also highlights external bottlenecks: electricity supply, capital intensity, and governance frameworks that could pace or distort AI's trajectory.