Anthropic and Accenture commit $1B+ to embedded third-party evaluation

Your AI and this story

Key facts

• Announced Sept 18, 2026
• At least $1B over five years in evaluation capacity
• Embedded evaluators get employee-level access to observe training and development decisions
• Non-exclusive; Anthropic will also work with METR and others; long-term it favors pooled or government funding
• State of the field (CNBC, Oct 11): METR said in August it had raised ~$71M in commitments over six months (vs $13.6M total 2024 contributions) and has fewer than 50 full-time staff; for-profit Vals AI grew from 8 to ~30 staff and raised $40M in August. Anthropic is in talks with METR and other nonprofits that plan to self-fund pilots of ‘elements’ of embedded evaluation
• The AI Evaluator Forum (incl. METR and AVERI) published a public letter in September, ‘Minimum Conditions for Embedding Evaluators’: transparency, protection from retaliation and access equivalent to the labs’ ‘own highly privileged employees’; embedded evaluation should complement, not replace, broader oversight (CNBC)
• OpenAI (Oct 9) said it is ‘actively finalizing contracts with third-party safety assessors’ and will announce details in the coming weeks, building on work with METR and Redwood Research; fired researchers Mikita Balesni and Tomek Korbak say they were dismissed over how they communicated with third-party evaluators, which OpenAI disputes (CNBC)
• Critics on funding and independence: Brown’s Suresh Venkatasubramanian (‘the problem, as always, is money’), Wharton’s Kevin Werbach (ecosystem ‘not robust enough’); Fathom CEO Andrew Freedman argues for government-licensed Independent Verification Organizations, as in the FRONTIER Act and two California laws Newsom signed (CNBC)