The Future of Life Institute graded Anthropic, OpenAI, Google DeepMind, and six rivals across 37 safety indicators. Anthropic finished first with a 2.66 GPA, a C+. xAI, DeepSeek, and Mistral all failed outright, and the panel's sharpest criticism was reserved for how far the industry's military ambitions have traveled since 2024.

Seven researchers spent months reading through model cards, safety policies, and confidential company surveys, then graded nine of the most powerful AI companies on Earth the way a professor grades an exam. When the results published on July 7, the highest mark in the class was a C+.

That grade went to Anthropic, the lab that has spent years marketing itself as the safety-conscious alternative to OpenAI and Google. The report, the Future of Life Institute's Summer 2026 AI Safety Index, confirms that positioning only in relative terms. Anthropic still finished on top. In absolute terms, the same reviewers who ranked it first also wrote that the company pursued "questionable military engagements," and found that not one of the nine companies graded, Anthropic included, has come close to proving it can control an AI system that becomes smarter than its makers expect.

"AI companies are sprinting toward a cliff," Max Tegmark, the MIT physicist who co-founded the Future of Life Institute, said in the release accompanying the report. "Despite acknowledging the great risks of artificial superintelligence, they continue racing to build it."

A Panel of Seven Graded Nine Companies on 37 Indicators

The AI Safety Index is not a new project. The Future of Life Institute, an independent nonprofit focused on reducing catastrophic risks from emerging technology, first published it in 2024 and has issued several editions since, grading a growing roster of companies each time. The Summer 2026 edition covers nine companies, more than any prior round, evaluated across 37 indicators grouped into six domains: risk assessment, current harms, safety frameworks, existential safety, governance and accountability, and information sharing.

The grading panel included seven researchers with no financial stake in the companies they reviewed: David Krueger of the University of Montreal and Mila, Sharon Li of the University of Wisconsin-Madison, Tegan Maharaj of HEC Montréal, Sneha Revanur of the youth advocacy group Encode, Robert Trager of the Oxford Martin AI Governance Initiative, Yi Zeng of Renmin University of China's Beijing Institute of AI Safety and Governance, and Stuart Russell, the UC Berkeley professor whose textbook trained a generation of AI researchers and who now directs Berkeley's Center for Human-Compatible AI.

Their evidence base combined public material, research papers, model cards, and news coverage, with a targeted survey the institute sent to each company. Five of the nine responded. Alibaba, xAI, DeepSeek, and Mistral did not, according to Axios, which reported on the index the day it published.

Anthropic Won. Its Winning Score Was Barely Passing.

Grades run on the standard American GPA scale, from a 4.3 A+ down to a zero F. Here is how the nine companies scored.

Company	Overall Grade	Score	Rank Change Since Winter 2025
Anthropic	C+	2.66	Still 1st
OpenAI	C	2.28	Top 3, unchanged
Google DeepMind	C	2.01	Top 3, unchanged
Meta	D+	1.32	Up from 6th to 4th
Z.ai	D-	0.88	5th
Alibaba Cloud	D-	0.87	6th
xAI	F	0.65	Down from 4th to 7th
DeepSeek	F	0.47	8th
Mistral	F	0.33	9th, first index appearance

Anthropic led five of the six graded domains, earning a B- in current harms and safety frameworks and a B in governance and accountability, on the strength of what the panel called relatively strong transparency and a comparatively established safety framework. OpenAI's strongest showing came in risk assessment, where the panel credited a broader evaluation suite and heavier engagement with outside testers.

The failing grades spanned three continents. "Inadequate safety is a global problem, not a regional one," the report's executive summary states, noting that xAI, DeepSeek, and Mistral each represent a different region, the United States, China, and Europe, and each still finished last in its class.

Mistral's failure carries a particular irony. The European Union has positioned itself as the world's most aggressive AI regulator, yet its most prominent homegrown lab scored the lowest of any company measured, at 0.33.

No Company Has Solved the Hardest Problem on the Test

Every company's weakest domain was existential safety, the panel's measure of how prepared a company is for AI systems that could escape meaningful human control. No company scored above a C-. Most scored D or below.

The panel did not dismiss the industry's efforts outright. It credited Anthropic's constitutional classifiers, OpenAI's calls for new governance institutions, Google DeepMind's monitoring commitments, and Meta's loss-of-control provisions as real attempts. It judged all of them "entirely inadequate." Reviewers were especially skeptical of the two techniques the industry leans on most, interpretability research and chain-of-thought monitoring, on the grounds that, as the report puts it, "detection is not prevention."

Russell delivered a version of that same warning in person just two days before the index published, at a United Nations AI governance conference in Geneva. "These systems are blackmailing, deceiving, launching nuclear weapons in tests," he said, according to an account relayed by the advocacy group PauseAI. "These are big, flashing red warning lights and fire alarms. It's not 'this is decades away.' You can hear those alarms sounding now."

UN Secretary-General António Guterres, addressing the same gathering, put the stakes in blunter terms still: "We may be the last generation able to set the terms on which humanity and machines coexist."

The Industry Quietly Walked Back Its Own Red Lines

Two findings in the report describe an industry retreating from commitments it made voluntarily, not under any regulatory pressure.

The first concerns military work. From 2024 to 2026, Anthropic, OpenAI, Google DeepMind, and Meta each maintained policies barring their technology from weapons systems or lethal autonomous operations. All four have since carved out exceptions to pursue defense contracts, joining xAI and Mistral, which never adopted such bans. Tegmark did not mince words describing the shift to Axios: "Boy oh boy has that changed."

The report singles out Anthropic specifically. Despite the company's stated limits on domestic surveillance and autonomous weapons, the review panel wrote that Anthropic drew criticism for "questionable military engagements," pointing to a reported link between Claude and the February 28 strike on an elementary school in Minab, Iran, which killed an estimated 120 children on the first day of US operations against the country. Claude is one of several AI tools built into Maven, the Palantir-built targeting system US Central Command used during the campaign.

Amodei addressed the strike directly on Bloomberg's "The Circuit" in June. "We don't have access to, we don't know exactly how these models were used," he said, adding that the use case would not have violated Anthropic's policies and that military leaders make mistakes "even at the best of times."

The second retreat concerns the pause pledges several labs made years ago, promising to halt development unilaterally if their systems approached specific danger thresholds. The panel found that Anthropic, OpenAI, Google DeepMind, and Meta have all weakened or scrapped those commitments. Reviewers called the pattern "moving goalposts" and concluded it has "undermined safety frameworks across the board."

Mistral Says the Report Card Misreads Open Weights

Not every graded company accepted the results quietly. Mistral, which finished last, told Axios that the index's methodology penalizes companies that release open-weight models rather than locking their systems behind an API.

"Mistral's models are open weight, which means enterprises decide how they're fine-tuned and deployed and can build in the specific safety controls their context requires," the company said in a statement. "A handful of companies deciding, behind closed doors, what's safe for everyone else is a risk that we would also highlight. Open, independently scrutinized models are the check on that concentration of power."

It is a real tension in how the index is built. A company that publishes its weights hands control of deployment to whoever downloads them, which the index counts against it under current harms and safety frameworks, while a company that keeps everything proprietary can claim tighter control it may or may not actually exercise. Mistral's own approach to openness has won it enterprise customers precisely because businesses can inspect and modify what they run. Whether that transparency counts as a safety feature or a safety gap is exactly what the report and Mistral disagree about.

xAI, now folded into the newly public SpaceX following February's $1.25 trillion merger of Elon Musk's two companies, did not respond to the institute's survey at all, and neither did DeepSeek or Alibaba. Their grades rest entirely on public material the panel could find without their cooperation, which likely understates whatever internal safety work, if any, those companies are not disclosing.

The Grades Are Already Shaping Boardroom Conversations

Tegmark told Axios the survey is "already having an impact in internal discussions" inside the companies it grades, and pointed to a detail easy to miss: more than half the nine companies participated this cycle, up from a group that mostly ignored the institute's outreach when the index launched. "It shows that they care," he said.

That matters for a different reason to the people who actually build on these systems. Every company on this list, from Anthropic and OpenAI down to DeepSeek, ships APIs and open weights that data scientists and ML engineers integrate into production every day. A safety grade is not a benchmark score, and it will not tell an engineer which model writes better code. It is closer to a credit report: a record of how a company behaves when nobody is forcing it to disclose anything, assembled by people who have no reason to flatter it.

The Bottom Line

Strip away the letter grades and the report describes an industry moving in one direction on capability and struggling to hold ground on safety. Every major lab is shipping more powerful systems faster than in 2025. None of them, according to the researchers who study this full time, has gotten meaningfully closer to proving those systems can be controlled if something goes wrong.

The uncomfortable detail is not that xAI, DeepSeek, and Mistral failed. It is that the company held up as the industry's safety leader, graded by a panel that included some of the most respected names in AI governance, still only cleared a C+, and did so while facing its own questions about military contracts. Anthropic, OpenAI, and Google DeepMind are, by this measure, the best the industry has. The index's verdict is that the best the industry has is not close to good enough.

"Detection is not prevention," the panel wrote of the interpretability tools every major lab is counting on. Nine companies are betting billions that they will not need prevention in time.