Up Close
OpenAI turns legal AI into a governed product, while Anthropic makes internal oversight measurable
OpenAI launched Astra for Law, a legal configuration of GPT-6 Astra for selected U.S. law firms and legal-technology companies inside ChatGPT and Codex, with a U.S. legal search index, tailored legal instructions, controls for confidential client work and 26 partner-built plugins. This matters less as another model variant than as a shift in where legal AI is bought: the package suggests firms are being asked to adopt one governed working surface, with search, workflow links and confidentiality controls already bundled, instead of assembling those pieces themselves. That appears to widen the pressure on Microsoft and everyone else selling enterprise AI into regulated work, because the contest is no longer only over model quality but over whether a provider can make a profession-specific stack usable without handing governance back to the customer. In this context, a plugin is the connector that lets the model act inside another software tool rather than staying in a chat window.
Anthropic’s Institute published internal metrics showing that in August Claude led 26% of Anthropic’s own AI research and development work, with more than 90% of tasks involving the model, about 30,000 agents running at once and real-time monitoring blocking around 0.002% of actions. We read this as an attempt to turn AI oversight from a promise into an operating figure: not just that monitoring exists, but how much work is delegated, how much is stopped and where compute is split between capability and safety. That is likely to matter outside Anthropic because enterprises and regulators now need ways to compare supervised use, not just model benchmarks, as AI systems gain more autonomy inside real workflows. An agent is software that can take steps on its own toward a task, meaning the issue is no longer only what a model answers but what it is allowed to do while running.