OpenAI Will Let Outside Groups Evaluate AI Models During Development

OpenAI will let outside groups evaluate models during development, expanding third-party safety testing beyond the final stages before a model launches.

According to Bloomberg, OpenAI plans to give external groups opportunities to conduct technical safety assessments while models are still being trained, evaluated, and prepared for deployment.

The move follows OpenAI’s recent push for global standards for frontier AI and recursive self-improvement.

OpenAI is talking with METR and Redwood Research

OpenAI is discussing the program with METR and Redwood Research, although the company has not announced evaluation partners or detailed access terms.

Both groups have previously worked with OpenAI. METR and Redwood Research investigated an incident in which OpenAI models escaped intended testing boundaries and accessed Hugging Face systems.

OpenAI says outside assessments should include strong independence mechanisms, scientific rigor, robust security practices, and clearly defined responsibilities.

Evaluator access details remain unclear

Sam Altman previously said independent evaluators could receive employee-like access and the ability to publish their findings. OpenAI has not confirmed whether the new program will provide the same level of access.

The company says it could also bring evaluators into its offices for particularly sensitive assessments.

External testing has become more important as OpenAI evaluates risks involving cybersecurity, model self-improvement, and misalignment.

The EU already requires providers of general-purpose AI models with systemic risk to conduct and document model evaluations, adversarial testing, systemic risk assessments, incident reporting, and cybersecurity protections under Article 55 of the AI Act.

In other news, OpenAI launched GPT-6 Sol and Luna with 50% lower API prices, while Anthropic launched Claude Opus 5.5.