Since September 2025, Free/Pro/Max users can opt in to training. Opted-in data retained up to 5 years de-identified. Even opted-out users' data may be used if flagged for safety review or via thumbs up/down feedback. Only Enterprise, Team, API, and education tiers guarantee no training. This is still stronger than most competitors, but the original "never trains" claim no longer holds.

The Pentagon refusal cost Anthropic its government contracts and earned a 'supply chain risk' designation. This demonstrated willingness to accept material financial harm to maintain stated principles.

In February 2026, Anthropic's RSP v3.0 removed its hard commitment never to train models unless safety measures were guaranteed. Co-founder Jared Kaplan said the change was because 'competitors are blazing ahead.'

Anthropic settled Bartz v. Anthropic for $1.5 billion, the largest copyright settlement in US history, stemming from downloading 7+ million pirated books. A separate $3 billion music lawsuit from Universal Music Group, BMG, and Concord is pending.

Anthropic's heavy reliance on Amazon, Google, Microsoft, and Nvidia for funding and compute creates a structural dependency. Burning ~$3B/year in cash. Break-even not expected until 2028.

Models are closed-source. Published safety research is extensive, but the models themselves cannot be independently audited.