Product news: ChatGPT research access, self-hosted Claude Code and Gemini Classroom put institutions in the control loop

This product-news report tracks a shared institutional shift across three different AI ecosystems. OpenAI is offering selected academic researchers expanded access to GPT-5.6, ChatGPT Work and Codex. Anthropic has placed self-hosted Claude Code environments into public beta for organizations that need agent execution inside their own network. Google is bringing teacher-led Gemini, study-notebook and NotebookLM activities into Classroom with curriculum grounding and educator insight. The products serve research, software and teaching, but each makes the surrounding institution more responsible for access, data, workflow design and review.

OpenAI's July 29 announcement starts ChatGPT for Academic Researchers with 10,000 researchers and aims to reach 100,000 through 2027. Participants at eligible institutions can receive GPT-5.6 models, expanded deep research, higher limits, larger context windows, research skills and connectors across ChatGPT, ChatGPT Work and Codex. OpenAI says the workspaces have business-grade privacy protections and that data are not used to train its models by default. The offer can widen access to capable tools, but free access and vendor benchmarks do not establish research quality. Institutions still need policies for confidential data, reproducible analysis, authorship, citation checking and disclosure of AI contribution.

Anthropic's August 6 release addresses where agent work runs. Claude Code sessions can now execute on organization-provisioned infrastructure, close to internal services, toolchains and security controls, while users start them from web, mobile, desktop or a routine. Repository checkouts, build artifacts, secrets and created files remain on that infrastructure. A crucial boundary remains: prompts, responses and tool results are sent to Anthropic for inference, and transcripts are stored so sessions can continue across surfaces. Team and Enterprise organizations must operate fixed or on-demand runners, isolate sessions and maintain the environment. Self-hosting therefore changes the control surface; it does not make data-flow review unnecessary.

Google's education release supplies the learner-facing comparison. Teacher-led Guided Learning, study notebooks and NotebookLM activities are designed to be grounded in selected class materials and to give teachers insight into individual and class interaction. The connected Classroom app in Gemini can use assignments, grades and materials to help educators analyze progress and draft activities. Google says Workspace for Education data are not used to train its AI models, while availability depends on account, language, platform and rollout stage. These are product claims and planned capabilities, not independent evidence of improved learning.

For Hong Kong schools and universities, the common lesson is to evaluate the whole workflow rather than the model alone. A pilot should identify who may start an agent or learning activity, where code and student data move, what logs and artifacts remain available, how users challenge sources, and who approves consequential outputs. Research access can be paired with reproducibility records; self-hosted execution with explicit inference-data maps; and teacher-led tools with independent assessments of what students can do unaided. Institutional control is valuable only when it produces inspectable decisions, protected data and credible evidence of learning or research quality.