Anthropic put a co-founder at the Vatican microphone. OpenAI put a tax-agent item at the top of its news page. Google put agents across its I/O front door. That is the frontier-AI week in miniature: one lab occupying the moral stage, two rivals turning the agent into a product surface. [1] [3] [4]
Wednesday's paper argued that OpenAI and Google answered AI week with agents, not Magnifica. It also put Anthropic's public virtue beside the compute bill that tests frontier-AI virtue. Thursday's update is not a reversal. It is a clearer split. The public record now shows Anthropic continuing to widen the moral conversation while OpenAI and Google keep making the agent legible as work.
Anthropic's strongest public artifact remains Chris Olah's remarks on Pope Leo XIV's encyclical, "Magnifica humanitas." Olah said every frontier AI lab, Anthropic included, operates inside incentives that can conflict with doing the right thing: commercial viability, research pressure, geopolitics, pride and ambition. [1] It was not a product launch. It was a confession of institutional constraint, delivered in the register of discernment.
The companion Anthropic essay on widening the conversation gives that speech a program. The company says it has held dialogues with scholars, clergy, philosophers and ethicists from more than 15 religious and cross-cultural groups. [2] It frames the work around moral formation, Claude's constitution, model character and the question of how an AI system that interacts with millions of people should be good. [2] This is not empty public relations copy. It is an attempt to make moral vocabulary part of the development process.
The question is whether that vocabulary travels into the market at the same speed as agents do. OpenAI's own news page is now an index of practical deployment. Its newest listed item is "Building self-improving tax agents with Codex," dated May 27. The surrounding entries sell enterprise coding agents, Codex partnerships, personal finance in ChatGPT, remote Codex work and a Windows sandbox. [3] The nouns are not conscience, flourishing or formation. They are Codex, sandbox, enterprise and agent.
Google is even less ambiguous. Its Innovation and AI page leads with I/O 2026 and the "agentic Gemini era," then points readers to Managed Agents in the Gemini API. [4] On the same page, Google directs developers to the product announcement that defines agents as files, runs them in secure cloud sandboxes and lets teams build with instructions, skills and data. [4]
The divergence matters because frontier labs no longer compete only on model benchmarks. They compete on public grammar. Anthropic wants to be heard as the lab that can say the incentives are dangerous. OpenAI wants Codex to look like a governed worker in enterprise software. Google wants Gemini agents to look like infrastructure, consumer assistance and developer tooling in one stack. [3] [4]
X sees this pattern faster than institutions want to name it. The platform frame is blunt: ship products, then talk ethics, unless the ethics talk itself is the product. This article leaves x_posts empty because no verified status URL is required by the assignment and no social post is needed to prove the divergence. The documents make the point. Anthropic's public language is thick with critics, conscience, religious communities and moral voices. [1] [2] OpenAI and Google are thick with agents, sandboxes, developer surfaces and AI products. [3] [4]
There is a charitable reading. Safety language can become real only if it changes product behavior. Agent products can be responsible only if they make control, auditing and failure boundaries concrete. A sandbox is a moral claim in engineering dress. A constitution is a product claim in moral dress. The frontier problem is that each side of that sentence is easier to market than to verify.
Olah's Vatican remarks are useful because they refuse the simplest corporate story. He said outside critics are needed because builders see from inside the incentives. [1] Anthropic's widening-conversation essay says the company hopes these dialogues will inform Claude's constitution, values and evaluations. [2] That is the right standard. The next receipt should not be another panel. It should be a product change or evaluation choice that can be traced to those conversations.
For OpenAI and Google, the test runs in the other direction. They have already made the agent concrete. The question is whether their agent infrastructure carries moral commitments beyond permissions, sandboxes and governance slogans. The market will buy work before it buys wisdom. That is why Anthropic's microphone matters, and why it is not enough.
-- ANNA WEBER, Berlin