AI is going vertical
Frontier AI firms are rolling out narrow, permissioned models aimed at specialists rather than the mass market. OpenAI introduced a biology‑tuned model called GPT‑Rosalind and a restricted cybersecurity variant called GPT‑5.4‑Cyber, while Anthropic released Claude Opus 4.7 with benchmark gains and updated safety framing. (venturebeat.com) (mashable.com).
The biggest artificial intelligence labs are starting to ship more restricted models for narrower jobs, not just bigger chatbots for everyone. OpenAI rolled out GPT‑Rosalind for life sciences on April 16, and two days earlier introduced GPT‑5.4‑Cyber for vetted defenders. (openai.com 1) (openai.com 2) GPT‑Rosalind is a research-preview model built for biology, drug discovery, and translational medicine, with stronger performance on tasks involving molecules, proteins, genes, pathways, and disease biology, OpenAI said. The company said qualified customers can use it in ChatGPT, Codex, and the application programming interface, and that a Codex plugin connects to more than 50 scientific tools and data sources. (openai.com) OpenAI said GPT‑5.4‑Cyber is a “cyber-permissive” variant of GPT‑5.4 for defensive work such as finding and fixing vulnerabilities. Access runs through its Trusted Access for Cyber program, which OpenAI said it is expanding to thousands of verified individual defenders and hundreds of teams. (openai.com) Anthropic took a parallel step on April 16 with Claude Opus 4.7, a generally available model that it said improves on Opus 4.6 in advanced software engineering, long-running tasks, instruction-following, and image understanding. Anthropic said Opus 4.7 is available across Claude products, its application programming interface, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. (anthropic.com) The common idea is specialization with gates. Instead of one model handling every user and every risk level, the companies are separating biology and cybersecurity work into products with screening, identity checks, or limited release terms. (openai.com 1) (openai.com 2) (anthropic.com) That approach lines up with the fields involved. Biology models can help sort papers, connect lab data, and propose experiments, but the same systems raise biosecurity concerns; cybersecurity models can speed up patching and red-teaming, but they can also lower the cost of misuse. (openai.com 1) (openai.com 2) (anthropic.com) OpenAI has been moving toward this structure for months. In June 2025, it published a biology safety post saying it was layering safeguards as model capability increased, and in February 2026 it launched Codex Security before expanding cyber access again this week. (openai.com) (openai.com) Anthropic is drawing a similar line between what is broadly sold and what is held back. In its Opus 4.7 announcement, the company said the model is less broadly capable than Claude Mythos Preview and that it reduced Opus 4.7’s cyber abilities during training while testing automatic blocking for high-risk requests. (anthropic.com) The business pitch is also getting more specific. OpenAI named Amgen, Moderna, the Allen Institute, and Thermo Fisher Scientific as GPT‑Rosalind users, while Anthropic kept Opus 4.7 pricing at $5 per million input tokens and $25 per million output tokens to make the upgrade look like a model swap, not a new product tier. (openai.com) (anthropic.com) For now, the shift is not away from frontier models. It is toward frontier models wrapped in permissions, industry workflows, and narrower claims about who should use them and for what. (openai.com) (openai.com) (anthropic.com)
