Stanford's AI Index 2026 shows rapid progress, growing safety concerns, and declining public trust

Model performance on benchmarks like MMLU and GPQA has improved 18–32% year-over-year, with frontier models now exceeding 85% accuracy. Safety evaluations, however, show a 12% drop in controllability scores across GPT-4, Claude 3, and Haiku, while public trust fell to 39% in the US.

Most teams still treat safety evals as a compliance checkbox, running them post-deployment in sandboxed environments. That’s backward. When Opus fails 41% of adversarial refusal tests, deploying it behind a RAG layer without real-time guardrails means you’re shipping a vulnerability. Running Opus for simple classification is just burning money.

Large-scale agent teams at fintech and healthcare startups must switch to continuous red-teaming with automated probes in staging. Small teams using Haiku for internal tools can ignore this—latency and cost still dominate their constraints.