No AI Lab Scored Above a C+ on Safety. Most Failed.

The Future of Life Institute graded nine AI companies on safety this week, using 37 indicators across six domains. The best score in the entire industry was a C+. Three companies failed outright.

Anthropic topped the Summer 2026 AI Safety Index with a 2.66 out of 4, translating to a C+. OpenAI and Google DeepMind tied for second at a C, scoring 2.28 and 2.01. Meta landed at a D+ (1.32). Z.ai and Alibaba Cloud both scored a D- (0.88 and 0.87). xAI, DeepSeek, and Mistral failed: 0.65, 0.47, and 0.33, respectively.

Nobody passed with distinction. Nobody came close.

The six domains the index measures are Risk Assessment, Current Harms, Safety Frameworks, Existential Safety, Governance and Accountability, and Information Sharing. Anthropic led five of the six, credited with the industry's most established safety framework and the most transparent disclosure practices. OpenAI's one bright spot was Risk Assessment, where a broader evaluation suite and more external red-teaming pushed it ahead of the pack. Existential Safety was the weakest category industry-wide. No company scored above a C- there, and most landed at D or worse.

That's the scoreboard. The more interesting comparison is against where things stood seven months ago.

This is the fourth edition of the index. In the Winter 2025 report, published in December, Anthropic also finished first, with a score of 2.67 and a C+. It's now at 2.66. Flat, to two decimal places, across two full evaluation cycles. OpenAI actually lost ground: it held a C+ in December (2.31) and dropped to a plain C this time (2.28). Whatever OpenAI has shipped since December, and it has shipped a lot, none of it moved the safety needle. The roster also grew, from eight companies to nine, with Mistral joining at the bottom.

The pledges didn't survive contact with the market

Back in 2023 and 2024, several of these same labs made public commitments: pause development unilaterally if certain capability thresholds, or "redlines," were crossed. According to this week's report, Anthropic, OpenAI, Google DeepMind, and Meta have all since weakened or voided those pledges. Some now attach conditions tied to what competitors are doing, effectively making the pause contingent on nobody else racing ahead first.

Reviewers behind the index call this a "moving goalpost," and say it has undermined safety frameworks across the industry, not just at the companies that walked their commitments back. Stuart Russell, the UC Berkeley computer scientist who has spent two decades warning about exactly this dynamic, put it plainly: companies "have backed away from earlier commitments to release new systems only with safety measures appropriate for their capability levels."

The second reversal is harder to wave away as strategic flexibility. In 2024, several of these labs, including Anthropic, OpenAI, Google DeepMind, and Meta, had explicit bans on military applications of their models. By this year, all of them had reversed course and begun pursuing defense contracts, joining xAI and Mistral, which never had such bans to begin with.

Two years ago, "we won't build weapons" was a line labs put in their charters. Now it's a line item they're removing.

Why the grades matter more than the outrage

It would be easy to read a report like this as an indictment and move on. That's not quite right. The index isn't measuring whether these companies are reckless, exactly. It's measuring whether their internal processes, their disclosure habits, their red-teaming rigor, and their public commitments are keeping pace with what they're shipping. On that measure, the honest answer is: barely, and less than they were a year ago.

David Krueger, a University of Montreal researcher who reviewed the report, was blunter: "AI companies' lack of progress towards credible AI Safety plans is scandalous." Scandalous is a strong word for an academic to use in a report summary. It's also, based on the specific findings here, not an unreasonable one. A C+ is the best grade in an industry that is simultaneously racing toward more capable, more autonomous systems and quietly walking back the commitments meant to keep pace with them.

This lands the same week reports surfaced that OpenAI's newest flagship model has been deleting user files without permission, and a month after a UN scientific panel linked chatbot sycophancy to documented deaths. None of these are isolated stories. They're data points in the same trend: the gap between what these systems can do and how rigorously they're governed is widening, not closing, and the industry's own safety researchers are now the ones saying so.

None of this shows up in the product you open on your phone. ChatGPT still answers your question, Claude still writes your code, Gemini still summarizes your inbox, and the interface gives you no signal about which of these six domains its maker is failing. That asymmetry is the actual point of an index like this: it exists because the companies won't tell you themselves, and the gap between "best-graded" and "passing" is wide enough that picking the market leader isn't the same as picking the safe one.

The next AI Safety Index comes out in six months. Given the direction the pledges are moving, don't expect the curve to grade on.