Nine AI labs took a safety exam. The best grade was a C+

The Future of Life Institute published its Summer 2026 AI Safety Index in July, grading nine frontier AI companies on safety practice. Nobody passed impressively. Anthropic earned the highest overall grade at C+ (2.66 on the index's 4-point scale), followed by OpenAI at C (2.28) and Google DeepMind at C (2.01). Meta received a D+, Z.ai and Alibaba Cloud each a D−, and three companies — xAI, DeepSeek and Mistral — received failing grades.

The grades come from a seven-member independent review panel that includes Stuart Russell of UC Berkeley, David Krueger of the University of Montreal and Robert Trager of Oxford, scoring each company across six domains: risk assessment, current harms, safety frameworks, existential safety, governance and accountability, and information sharing.

The finding under the grades

Letter grades make headlines, but the report's sharpest finding is about promises. Several leading developers had previously published safety frameworks pledging to pause development or restrict releases on their own if systems approached defined risk thresholds. The panel found that industry leaders have, in its words, “weakened or voided pledges to pause unilaterally if redlines are approached” — a pattern it characterizes as moving the goalposts, and one it says has undermined safety frameworks across the board.

That reframes what a C+ means. These are not grades against an abstract ideal; much of what the index measures is whether companies keep standards they set for themselves. The industry's best performer leads five of six domains — Anthropic takes a B+ in information sharing and a B in governance and accountability, while OpenAI leads risk assessment — and still averages out to a C+. The panel's implication is uncomfortable in both directions: the leaders are mediocre by their own stated standards, and the gap between the leaders and the failing tier is enormous at exactly the moment these systems are being wired into security tooling, health workflows and autonomous agents.

What is not established

An index is a lens, not an inspection. The grades rest on public documentation, company questionnaire responses and expert judgment — not on regulatory audit or internal access, so a company that discloses little can be graded harshly for opacity rather than for practice, and a polished framework can score well on paper regardless of how it is applied. The panel is transparent about its methodology, but reasonable people weight these domains differently, and no government has adopted this scale. The index's real function is comparative and longitudinal: the same panel, the same rubric, every edition — which is precisely what makes the weakened-pledge finding hard to dismiss.