The AI Safety Index 2026: Nobody Earns an A, and Labs Are Walking Back Their Commitments

The Future of Life Institute publishes its AI Safety Index twice a year, and the Summer 2026 edition carries a verdict that is difficult to spin: not one of the nine largest AI laboratories in the world — companies collectively worth trillions of dollars and deploying systems used by hundreds of millions of people daily — has earned a passing grade by the evaluators’ standards. The highest score, a C+, went to Anthropic. Four labs received outright failing grades.

The index is not a fringe document. It is compiled through a structured assessment process involving independent researchers, drawing on public commitments, disclosures, incident reports, and laboratory conduct across six evaluation dimensions. Its methodology is contestable, as all such frameworks are, but it is taken seriously by the policy community and has become a reference point in regulatory discussions from Washington to Brussels.

The Scorecard

The nine companies evaluated were Anthropic, OpenAI, Google DeepMind, Meta, Z.ai (the rebranded Zhipu AI), Alibaba Cloud, xAI (Elon Musk’s AI company), DeepSeek, and Mistral. The grades:

• Anthropic: C+
• OpenAI: C
• Google DeepMind: C
• Meta: D+
• Z.ai: D-
• Alibaba Cloud: D-
• xAI: F
• DeepSeek: F
• Mistral: F

The six evaluation dimensions were: risk assessment, current harms mitigation, safety frameworks, existential safety, governance and accountability, and transparency and communication. No company performed well across all six categories. Anthropic’s relative strength was in safety frameworks and existential safety; its relative weakness was in current harms mitigation and transparency. OpenAI’s profile was broadly similar. The Chinese and European companies that failed did so primarily on governance, transparency, and safety framework dimensions.

The Commitment Reversal Problem

The most substantive finding — and the one generating the most commentary — is not the grades themselves but the pattern of commitment reversals the report documents.

From 2023 onward, major AI labs made public statements about categories of work they would not pursue: training systems with insufficient safety guarantees, enabling military applications, deploying systems that could cause catastrophic harm without adequate safeguards. Many of these commitments were made voluntarily in advance of regulation, and some were incorporated into the Seoul AI Safety Statement and other intergovernmental agreements.

The 2026 index documents that nearly all of these commitments have been quietly weakened or reversed. Anthropic, OpenAI, Google DeepMind, and Meta — companies that had all previously stated policies against military applications — have since shifted to actively seeking defense partnerships. In the current funding environment, where US and European defense agencies are among the most significant AI procurement buyers, the commercial incentives are obvious.

The most specific and consequential reversal the report highlights is Anthropic’s. In February 2026, the company withdrew its previous pledge not to train AI systems unless it could guarantee in advance that its safety measures were sufficient. The pledge, known internally as the “Constitutional AI training commitment,” had been a cornerstone of Anthropic’s differentiation from competitors. Its withdrawal was not announced proactively; it was discovered by researchers monitoring the company’s published policies.

Anthropic’s representatives have argued that the pledge was operationally impractical as models became more capable — that the bar it set was impossible to meet without abandoning frontier AI research entirely. Safety advocates have argued this proves the pledge was never meaningful, or that it reveals the company’s commercial pressures are now overriding its safety culture.

What C+ Actually Means

It is worth being specific about what “best-in-class” looks like in this assessment, because it is easy to interpret Anthropic’s C+ as an endorsement when it is not.

The C+ reflects genuine relative strengths: Anthropic has more mature safety evaluation frameworks than competitors, publishes more detailed policy documents, maintains the most explicit model card disclosures, and has invested more heavily in interpretability research than any publicly comparable organization. These are real differences that matter at the margin.

What the C+ also reflects is a fundamental gap between stated aspirations and demonstrated accountability mechanisms. Anthropic, like its competitors, has no independent board with authority over safety-critical decisions, no external audit regime with meaningful enforcement power, and no legal obligation to publish training incident reports. Its safety evaluations are conducted largely by its own safety team — a team whose senior members are employed by the company being evaluated. The circularity of this arrangement is not unique to Anthropic; it applies to the entire sector. But it means the evaluators have limited ability to verify what companies claim about their own practices.

The Failing Labs

The F grades for xAI, DeepSeek, and Mistral reflect structural rather than merely technical deficiencies. All three operate with minimal public safety documentation. None publishes anything resembling an evaluation framework for potential harms before deployment. DeepSeek, the Chinese lab that sent shockwaves through Western AI markets with its February 2025 V3 release, has consistently declined to engage with international safety norms discussions.

Mistral’s failure is perhaps the most striking given its European origins and the regulatory environment in which it operates. The EU AI Act imposes compliance obligations on providers of high-risk AI systems, and Mistral’s flagship models fall within scope. Yet the index finds Mistral’s safety documentation and governance structures to be among the thinnest of any company evaluated. European regulators have noted this gap privately; whether the AI Act’s enforcement mechanisms will close it remains to be seen.

The Governance Gap

The index’s broader argument — implicit in the scorecards but explicit in its analysis sections — is that the current self-regulatory model for AI safety is failing. The labs with the best safety cultures and most sophisticated evaluation frameworks, Anthropic and OpenAI, still earn grades that would be academic failures for any serious purpose. The labs with the least developed safety cultures operate with near-impunity in their home markets.

What would meaningful governance look like? The index does not prescribe specific policies, but its framework implies several requirements: independent third-party audits with genuine access to model weights, training procedures, and incident records; mandatory disclosure of adverse findings; legal frameworks that create liability for documented safety failures; and international coordination to prevent regulatory arbitrage.

None of these mechanisms currently exist at meaningful scale. The EU AI Act creates some disclosure requirements for general-purpose AI models, but enforcement has been slow and the accountability mechanisms remain largely untested. The US has relied on voluntary commitments — which this index documents have not held.

The Industry’s Response

Major labs have responded to the index with a mix of engagement and deflection. Anthropic acknowledged the C+ without challenging the methodology, calling it “a useful external perspective” while noting that it disagrees with specific criteria weightings. OpenAI issued a similar statement emphasizing its ongoing safety work. Meta did not respond publicly. The labs that received failing grades — xAI, DeepSeek, Mistral — did not comment.

The silence from the F-graded labs is itself informative. In a sector that has learned to manage reputational risk with sophisticated communications strategies, choosing not to engage with a published safety assessment is a signal about how those organizations prioritize public accountability.

Why This Matters Right Now

The timing of this index is not incidental. AI systems are more capable than at any previous assessment, more widely deployed in consequential settings — medical, legal, financial, educational, security — and more deeply embedded in enterprise and government workflows. The safety stakes have increased as the capabilities have advanced.

The trajectory documented by consecutive index editions is the most concerning finding: not just that labs are falling short of their own standards, but that their stated standards are declining. If safety commitments erode fastest precisely when AI systems become powerful enough for those commitments to matter most, the governance frameworks that were supposed to hold through this transition are instead giving way under the commercial pressure of the AI boom.

That is the finding that deserves the most attention, regardless of how one assigns letter grades.