The Future of Life Institute published its AI safety index for summer 2026 on July 19, reassessing nine major AI providers. Anthropic leads the field with a grade of C+ and 2.66 out of a possible four points, while Mistral from France is at the bottom of the table with 0.33 points. None of the companies evaluated achieved a better grade than C+.
Seven experts assess 37 safety indicators
An independent panel of seven researchers – including Stuart Russell from the University of California, Berkeley, and David Krueger from the University of Montreal – assigned grades based on 37 indicators in six areas. These include risk analysis, current harms, safety frameworks, existential risks, corporate governance, and information sharing. The basis was publicly available model cards, research papers, and targeted company surveys up to June 3. The panel evaluated only policies, governance structures, and published frameworks, not the actual performance of the delivered products. Following Anthropic in the ranking are OpenAI (C) and Google DeepMind (C), with companies Meta (D+), Z.ai, and Alibaba Cloud (both D-) as well as xAI, DeepSeek, and Mistral receiving an F. Thus, one company each from the USA, China, and Europe failed. In the sub-area of existential safety, none of the nine companies exceeded a C-, with most remaining at D or worse. The panel assessed the methods used there, such as interpretability tools, as mere post-recognition of problems rather than as prevention.
Top companies relax previous safety promises
The report criticizes especially the four top-ranked companies Anthropic, OpenAI, Google DeepMind, and Meta. They have weakened previous commitments to automatic development pauses for risky models and linked them to the behavior of competitors instead of adhering to them independently. Between 2024 and 2026, the same companies also lifted previous restrictions on military applications and are now actively pursuing defense contracts. Stuart Russell stated according to Tech Times that the race for capabilities has intensified: companies are retreating from safety measures and planning releases despite recognizable safety gaps. Panel member David Krueger added that the companies lack robust safety plans; executives hinted at pauses but did not communicate urgency or concrete preparations. Compared to the previous index from winter 2025, OpenAI declined from C+ to C, while Meta improved from a D to D+. Both xAI and DeepSeek dropped from a D to an F after previously being in the midfield. Only Anthropic made noticeable improvements in areas such as governance and information sharing, achieving the best scores in the entire field there.
Mistral rejects criticism of the methodology
Mistral counters the poor rating according to a report from the French portal Cryptoast with the argument that the methodology is tailored to closed models and systematically disadvantages open, freely downloadable systems. In such models, the responsibility for calibration lies more with the users than with the manufacturer, the company argues. Z.ai, which also predominantly publishes open models, performed poorly in the index – a pattern that supports Mistral’s objection. The Future of Life Institute stated that it had contacted Mistral multiple times; unlike most other evaluated companies, Mistral did not respond to the questionnaire. Mistral also points to its compliance with the European AI regulation as evidence of its safety level.
It will be crucial whether the poor grades have consequences for public tenders or corporate clients, who are increasingly relying on supplier audits. For Mistral, more than just its image is at stake: as the only European provider in the test, the rating provides new ammunition for the debate on the implementation of the EU AI regulation, whose next obligations will come into effect in August 2026.
