Two companies have each committed at least $1B over five years to build out independent evaluation of frontier models.
Anthropic announced the arrangement on September 18 with Accenture, whose specialist AI business Faculty will run the work. The scope covers red-teaming, alignment assessment, safeguard testing and general model evaluation.
The unusual part is where the evaluators sit. Traditional third parties assess a lab from outside. Embedded evaluators work inside it, with access comparable to an employee, which Anthropic says lets them watch a model develop during training, trace the decisions that shape deployment, and talk directly with staff. In that position they can check whether safety commitments hold, flag incidents and give the public a fuller account of risk.
Anthropic frames the arrangement as adding verifiability without moving accountability: model safety, it says, remains the company’s own responsibility. Plenty of operating detail is still undecided.
The commitment traces to a September essay by CEO Dario Amodei titled We Must Pace the Frontier, which named three steps – embedded evaluators, coordination between frontier labs in democratic countries, and global agreements. Anthropic adopted the first unilaterally and asked governments to make competitors follow.
