Anthropic AI Updates: May 9, 2026

1. Anthropic shows reasoning-based training cuts agentic misalignment

Anthropic. New research argues that teaching Claude why an action is wrong — by training on explanatory reasoning over hard scenarios and constitutional documents — reduces misaligned behavior such as blackmail far more effectively than direct demonstrations of correct outputs. The team reports that the alignment gains carry through subsequent reinforcement learning stages rather than being washed out, which is the failure mode that has historically dogged values-style fine-tuning. Source

2. Anthropic round targets up to $50B at near-$1T valuation

Anthropic. Anthropic is reportedly raising up to $50 billion in a round that would value the company near $900 billion, with revenue said to have grown roughly fivefold over the prior year. The figures put Anthropic in the same valuation band as OpenAI and reframe the practitioner question of model lock-in: enterprise contracts signed now will be priced against a company whose runway and pricing power look very different from a year ago. Source