Welcome to the agent platform research briefing for Tuesday, July 28th, 2026.

**OpenAI ChatGPT Health expands to all U.S. adults** — OpenAI has rolled out its Health feature in ChatGPT to all American users 18 and older across Free, Go, Plus, and Pro tiers. Launched in limited preview on July 23rd, the full expansion allows users to connect Apple Health and compatible medical record providers directly into ChatGPT conversations. The AI can now contextualize lab results, track medication changes, and support health conversations — handling 300 million weekly health queries with access to real medical data. Notably, the feature operates outside HIPAA coverage, raising regulatory questions. Available on web and iOS versions of ChatGPT 1.2026.188 or later. Roll-out may take a few weeks to reach all eligible users.

**Google confirms Gemini 4 training underway** — While Gemini 3.5 Pro remains months behind schedule and stuck in delays, CEO Sundar Pichai has confirmed Google is already training Gemini 4 in early stages. Pichai told 9to5Google that the next frontier generation will require "a larger base model to compete at that level," calling the training run "very ambitious." This marks Google's response to falling behind Anthropic and OpenAI on agentic coding benchmarks. Meanwhile, Google DeepMind has launched three models simultaneously: Gemini 3.6 Flash with 17% fewer tokens needed for comparable results, Gemini 3.5 Flash-Lite for cost-effective agent workloads, and Gemini 3.5 Flash Cyber — a cybersecurity-specialized model that found 55 confirmed unique vulnerabilities in Google's V8 JavaScript engine under fixed testing, outperforming Claude Opus 4.6's 36. Flash Cyber is piloting with governments and trusted partners only.

**Anthropic ships Claude Opus 5 with API innovations** — Anthropic's Claude Opus 5, released July 24th, is now the default model on Claude Max and strongest on Claude Pro, priced at $5 input and $25 output per million tokens — half the cost of Fable 5 while matching its performance on key benchmarks. But the bigger news came in a follow-up release: two beta API features that reshape how developers build agents. Mid-conversation tool changes let agents discover and connect new tools — databases, browsers, code runners — after a task has already started. And automatic fallbacks with a new "default" mode redirect safety-flagged requests to recommended alternative models instead of hard-blocking them, organized by refusal category. This is the first time a frontier lab has baked fallback intelligence directly into the API rather than leaving developers to manage their own fallback chains.

**OpenAI enterprise analytics and spend controls** — OpenAI has launched new usage analytics and updated spend controls for ChatGPT Enterprise, designed to give clients greater visibility and management over their AI deployments. The features include per-user utilization dashboards, cost attribution by team or project, and configurable spending caps. This comes as major employers including Tesla, Uber, Meta, and Amazon have all instituted AI spending limits after engineers burned through token budgets. With ChatGPT now handling 300 million weekly health queries alone, usage visibility is becoming a critical requirement for enterprise buyers.

That's the briefing for July 28th. Four stories shaping the agent platform this morning.