If the world's leading artificial intelligence companies were students, none of them would be making the honour roll. The Future of Life Institute's Summer 2026 AI Safety Index, released this month, graded nine frontier AI developers on their safety practices - and the best performer in the entire industry managed only a C+. At the very moment AI systems are being wired into cybersecurity operations, healthcare workflows and autonomous agents that act on users' behalf, the independent scorecard suggests even the most safety-conscious lab is doing a middling job by the standards it sets for itself.

The report card, company by company

Anthropic, the maker of the Claude models, finished first with a C+ - the index's highest grade - leading five of the report's six assessment domains on the strength of comparatively strong transparency, an established safety framework, technical safety research and governance structures. OpenAI and Google DeepMind followed with a C each, with OpenAI singled out as the leader in the risk-assessment domain thanks to a broader evaluation suite and more diverse engagement with external testers. Meta earned a D+, though the report notes it improved its overall ranking from sixth to fourth. At the bottom, xAI - which slid from fourth place to seventh - joined DeepSeek and Mistral in effectively failing the assessment.

  • Anthropic: C+ (highest grade; leads five of six domains)
  • OpenAI: C (leads the Risk Assessment domain)
  • Google DeepMind: C
  • Meta: D+ (improved from 6th to 4th place overall)
  • xAI, DeepSeek, Mistral: effectively failing grades (xAI fell from 4th to 7th)

How the index is put together

The Future of Life Institute is a nonprofit that has been publishing these periodic safety scorecards with a panel of independent AI experts as reviewers. The Summer 2026 edition combined publicly available materials - model cards, safety frameworks, published research, policy documents - with responses to a targeted survey sent to the companies themselves, with evidence collected up to June 3, 2026. Companies are graded across six domains covering areas such as risk assessment, transparency, governance, and safety frameworks, and the domain scores roll up into the headline letter grade.

The finding that worried reviewers most

Beyond the individual grades, two systemic findings stand out. The first is that existential safety - preparedness for the most severe, large-scale risks from increasingly capable systems - was the weakest domain across the entire industry, with no company scoring above a C- in it. The second is a pattern the reviewers described as moving the goalposts: Anthropic, OpenAI, Google DeepMind and Meta have all weakened or voided earlier pledges to unilaterally pause development if certain risk red lines were approached, in some cases making those commitments contingent on what competitors do. In the reviewers' assessment, that shift has undermined safety frameworks across the board - a commitment that dissolves under competitive pressure was never much of a commitment at all.

Why this matters to ordinary users

It is tempting to file an AI safety report card under industry inside-baseball, but the timing argues otherwise. AI assistants are drafting emails, summarising medical information, screening job applications and executing multi-step tasks with real-world consequences - and adoption in India is among the fastest anywhere, with recent industry surveys suggesting a large majority of Indian users already touch generative AI features and most device buyers now weigh AI capability in purchase decisions. When the companies building these systems collectively score between C+ and F on independent safety assessment, the gap between deployment speed and safety maturity becomes everyone's problem, not just the industry's.

The takeaway

The index's most useful contribution may be less the letter grades than the trajectory they reveal. Competition among frontier labs has visibly intensified, and the report documents safety pledges bending under that pressure rather than holding firm. For policymakers weighing AI regulation - in India, Europe and the United States alike - the message is uncomfortable but clear: self-regulation is producing C students at best. For users, the practical advice is unchanged but sharper: treat AI outputs in high-stakes domains with scepticism, prefer providers that publish their safety practices, and remember that a C+ is currently as good as it gets.