Welcome back to Tech Pulse β the daily, no-fluff rundown of what actually happened in tech and why it matters. Today (June 4, 2026): OpenAI quietly closes the GPT-4 chapter, the whole industry stops arguing about models and starts fighting over where AI agents run, Nvidia's next data-center engine moves into production, and astronomers hand the public the clearest map of the universe's skeleton yet.
π€ AI models & tools
OpenAI is closing the GPT-4 era. OpenAI confirmed it's retiring GPT-4.5 β with the model leaving ChatGPT on June 27, 2026 after a sunset window β alongside the older o3 reasoning model, while pushing an upgrade to GPT-5.5 Instant (gHacks, BleepingComputer). GPT-5.5 has been the flagship since April; this is housekeeping, not a new launch. Why it matters: if you call GPT-4.5 or o3 through the API, you have a migration to plan β test your prompts on GPT-5/GPT-5.5 now rather than on the 27th. Pinned model names are convenient until the day they're switched off, which is one more reason the shift to usage-based AI billing deserves attention. The fun part: GPT-4 is the model that made "AI" a dinner-table word in 2023. Three years later it's being quietly walked out the back door.
**The real 2026 fight isn't the model β it's the runtime. Google made Gemini 3.5 Flash generally available and added Managed Agents to the Gemini API (its "Antigravity" agent can spin up a remote Linux box to plan, run code, and browse), with Google saying Flash beats its own Gemini 3.1 Pro on coding and agentic benchmarks (Google). At the same time, Microsoft and Nvidia are building Windows "agent" plumbing β new security and containment primitives plus an OpenShell** runtime so agents can act on your PC without running wild (Nvidia). Why it matters: everyone now assumes the models are good enough. The land grab has moved to the control plane β the runtime, sandbox, and permissions that decide what an agent is actually allowed to do. New to the term? Start with what AI agents actually are, then see why so much of this is shifting toward on-device AI.
π» Hardware & infrastructure
Nvidia's next big engine, Vera Rubin, is moving into production. At its GTC keynote, Nvidia said the Vera Rubin platform is entering full production, naming Samsung, SK hynix, and Micron as HBM4 memory suppliers, with first customer shipments targeted for this summer and broader availability in the second half of 2026 (SiliconANGLE, Nvidia). Nvidia says an NVL72 rack pairs 72 Rubin GPUs with 36 Vera CPUs, and pitches large efficiency gains over the current Blackwell generation β figures worth treating as vendor claims until independent benchmarks land. Nvidia also says the companion Vera CPU is in production with early adopters including OpenAI, Anthropic, and SpaceX. Why it matters: this is the data-center counterpart to the consumer "AI PC" push we covered in yesterday's Pulse β and the hardware bill behind every chatbot and agent you use. For the consumer side of the same story, see our AI PCs: hype vs. reality breakdown.
π¬ Science & inventions (the cool stuff)
Autism might not be one condition. Researchers combined fMRI brain scans from nearly 1,000 autistic people with 20 genetically engineered mouse models and found evidence for at least two biologically distinct subtypes: a "hypoconnectivity" pattern (brain regions talking less) tied to cell-signaling pathways, and a "hyperconnectivity" pattern (talking more) tied to immune-related pathways (ScienceDaily). The honest caveat: this is research, not a diagnostic test β and about 75% of participants didn't fit cleanly into either subtype. It's a clue about biology, not a new label for the clinic.
A possible "off switch" for exhausted cancer-fighting cells. CAR T-cell therapy can lose steam as the engineered immune cells get "exhausted." A team reported that a transcription factor called NFIL3 is a major driver of that burnout β and that removing it (via CRISPR) kept the cells active longer with stronger anti-tumor effects in animal studies (ScienceDaily). Why it matters: exhaustion is one of the biggest reasons cell therapies fade against solid tumors. (Animal-stage β promising, not a treatment yet.)
The clearest map yet of the universe's hidden skeleton. Astronomers used NASA's James Webb Space Telescope and the COSMOS-Web survey β more than 164,000 galaxies β to build the most detailed map ever of the cosmic web, the vast network of filaments connecting galaxies, traced back to when the universe was about a billion years old. The team publicly released the maps and the full catalogue (UC Riverside, Phys.org). The fun part: the "empty" space between galaxies isn't empty at all β it's strung together like cosmic cobwebs, and now anyone can download the data.
π The one-line recap
OpenAI retired GPT-4.5, the AI race moved from models to agent runtimes, Nvidia's Vera Rubin headed for production, and Webb mapped the cosmic web in record detail. For more, browse our AI News & Analysis and Future Tech topics.
Tech Pulse runs daily. Numbers and features change fast β we link primary sources so you can verify, and we flag company claims as company claims. See you tomorrow.
