Most of the time, AI companies operate in a fiercely competitive environment, treating one another as rivals locked in a race to dominate the cutting edge of technology. Yet in a surprising turn, OpenAI and Anthropic have revealed that they recently undertook a cooperative effort to evaluate the safety and alignment of each other’s publicly available AI systems—and they have shared the results of their analyses with the public.
The full reports are highly technical, diving deep into the mechanics of AI alignment, instruction adherence, and risk mitigation. For those following the evolving landscape of AI development, they are a fascinating read. Broadly, the findings highlighted some weaknesses in each company’s models while also offering valuable insights into improving safety testing protocols moving forward.
Anthropic focused its evaluation on OpenAI’s systems, testing for issues such as sycophancy, whistleblowing, self-preservation, enabling human misuse, and capabilities that could undermine AI safety evaluations and regulatory oversight. The company reported that OpenAI’s o3 and o4-mini models generally performed in line with Anthropic’s own AI systems, but concerns were raised about potential misuse of the GPT-4o and GPT-4.1 general-purpose models. Sycophancy—AI’s tendency to overly agree with user input—was flagged as an issue across most tested models, though the o3 model showed the least tendency toward this behavior.
It’s important to note that these tests did not include OpenAI’s latest release, GPT-5, which features a system called Safe Completions designed to protect users and the public from queries that could be potentially harmful. This comes in the wake of OpenAI facing its first wrongful death lawsuit, following a tragic incident where a teenager engaged with ChatGPT about suicide over several months before taking his own life. The case has intensified scrutiny on AI safety, particularly concerning vulnerable users like minors.
On the other side, OpenAI conducted its own evaluation of Anthropic’s Claude models, focusing on instruction hierarchy, jailbreaking, hallucinations, and the potential for malicious scheming. The Claude models performed well overall: they adhered closely to instructions and demonstrated a high refusal rate when prompted with queries that could lead to hallucinations or potentially inaccurate responses. This suggests that Anthropic’s systems are generally cautious when faced with ambiguous or risky prompts.
The fact that these companies are conducting joint assessments is notable, especially considering recent tensions between them. OpenAI was reportedly restricted from accessing Anthropic’s tools earlier this month after allegedly violating the company’s terms of service by using Claude models in the development of new GPT systems. Despite the competitive friction, this collaboration underscores a broader shift in the AI industry: as tools grow more powerful and integrated into daily life, safety and alignment are becoming priorities that transcend corporate rivalries.
With regulatory attention mounting and public concern over AI risks rising, efforts like these may set an important precedent. By testing each other’s models rigorously and transparently, OpenAI and Anthropic are taking steps toward establishing a new standard in responsible AI development—one that balances innovation with user safety and ethical accountability.