Third-party evaluators will audit Claude with office access: $2 billion over five years

The embedded evaluator team moves inside the company. They test models, conduct red-teaming, check alignment and safeguards. They are promised employee-level access: to observe training, see release decisions, and speak directly with those making them.

Publication without clearance. Evaluators can publish findings on risk levels, incidents, and practices without Anthropic editorial control. The company retains only the right to remove confidential data.

The assessed party pays. For now, Anthropic itself pays for Accenture’s work. The partnership is non-exclusive: Accenture may work with other developers, while Anthropic is negotiating with the nonprofit METR on a pilot funded by the organization itself.

On Accenture’s side, the work is led by Faculty — a British AI company that Accenture acquired on 16 March 2026 along with more than 400 AI specialists. The team is overseen by Marc Warner, Accenture CTO and Faculty CEO at the same time.

The industry did not expect a consulting company to be chosen: METR, Redwood Research, and Apollo Research had been tipped for the evaluator role. Some critics call the arrangement a way to dodge accountability; Anthropic responds that evaluators “do not reduce our responsibility.”

The move fulfills the first of three commitments in Dario Amodei’s 12 September 2026 essay “We Must Pace the Frontier” — permanent employee-level access for third-party evaluators.

Anthropic promises to name the next evaluators in the coming weeks.