OpenAI and Anthropic have jointly tested each other’s AI models to spot safety issues that might be missed by internal reviews.

The collaboration focused on hallucinations—when AI generates false information—and misalignment, where models do not behave as intended.

The tests were completed over the summer, before OpenAI launched GPT-5 and before Anthropic released its Claude Opus 4.1 update in August. Both firms are leading competitors in AI, with Anthropic founded by ex-OpenAI staff.

AI companies face increased scrutiny over product safety after reports of harmful outcomes, including a lawsuit against OpenAI alleging a teenager died by suicide after using its chatbot as a coach.

OpenAI described the effort as the first major cross-company exercise in AI safety testing.

🔗 Source: Bloomberg

🧠 Food for thought

1️⃣ Competitive dynamics in AI create paradoxical relationships between rivals

The OpenAI-Anthropic collaboration reveals the complex tensions shaping the AI industry, where companies simultaneously compete and cooperate on safety issues.

Earlier this year, Anthropic cut off OpenAI’s access to its Claude models after discovering OpenAI was using them for competitive benchmarking, violating Anthropic’s terms of service 2. Yet both companies still conducted mutual safety evaluations this summer, with Anthropic maintaining “limited access for benchmarking and safety evaluations” 2.

This selective collaboration pattern reflects how AI companies are navigating the balance between protecting competitive advantages while addressing shared industry-wide safety concerns.

OpenAI called this the “first major cross-lab exercise in safety and alignment testing” 1, suggesting that despite fierce competition—including ongoing pricing wars where OpenAI has reduced GPT-4 pricing by 3x and captured 80% of enterprise generative AI market share 3—safety issues create common ground for cooperation.

The collaboration is notable given that Anthropic was founded by former OpenAI employees and the companies represent different philosophical approaches: Anthropic emphasizes safety through “Constitutional AI” while OpenAI focuses on rapid progress and accessibility 4.

2️⃣ External safety evaluations reveal blind spots that internal testing misses

The joint evaluation exercise demonstrates how companies can miss critical safety issues in their own products, making third-party assessment valuable for risk management.

Anthropic’s evaluation raised concerns about potential misuse of OpenAI’s GPT-4o and GPT-4.1 models, while noting that the o3 model performed better 5. Meanwhile, OpenAI assessed Anthropic’s Claude models for instruction hierarchy, jailbreaking, hallucinations, and scheming, finding they generally performed well 5.

Recent OpenAI developments