Rosetta Intel

Daily developments from OpenAI / Anthropic / Google DeepMind · Window: past 24 hours (2026-06-02 → 2026-06-03)
🛡 = security-related

OpenAI

OpenAI frontier models + Codex GA on AWS (6/2)— enterprises can reach production faster through AWS-native security, governance, procurement, billing, and compliance. Another step in OpenAI's multi-cloud distribution. Releasebot
Model retirement schedule— GPT-4.5 exits ChatGPT on 6/27(30-day sunset); OpenAI o3 exits on 8/26(90-day sunset). OpenAI model release notes
Codex updates— macOS "Appshots" (a hotkey attaches an app window to a Codex thread); Goal mode fully rolled out(app / IDE extension / CLI). Response style and quality refinements for GPT-5.5 Instant. Releasebot
GPT-5.6 reportedly set for June— media reports continue; unconfirmed by OpenAI. Geeky Gadgets
Considering open-sourcing its scheduling software— compute lead Sachin Katti said the company is weighing open-sourcing its scheduling software, aiming to improve multi-vendor chip compatibility and challenge NVIDIA's dominance. GuruFocus

Confidential IPO filing (6/1)— immediately following the $65B Series H at a $965B valuation; annualized revenue reportedly around $47B(mostly from selling Claude subscriptions to enterprises and individuals). Seen as the first major public-market test of the AI boom. Washington Post· CBS News
Broad Claude outage (6/2)— elevated error rates across multiple Claude services, which Anthropic attributed to capacity constraints; resolved the same day. The National
Claude Mythos / Project Glasswing expands to about 150 companies— announced after the IPO filing, widening enterprise access to Mythos-class models; a continuation of the "AI finds bugs at scale" throughline (the 5/22 first monthly update remains the reference point: 10,000+ high/critical vulnerabilities in 30 days). Yahoo Finance· Glasswing monthly update
Billing changes from 6/15— the Claude Agent SDK, claude -p, Claude Code GitHub Actions, and third-party agents move off subscription allowances onto a separate monthly metered allowance (at full API pricing, no rollover). Heavy agent / CI users should assess costs ahead of time. Codersera analysis

GTIG/DeepMind — nation-state attackers using Gemini across "all phases" of attacks— groups linked to China, Russia, Iran, and North Korea are using Gemini for OSINT/target profiling, phishing lures, vulnerability research, scripting, and post-exploitation; the Honestcue malware dynamically generates and executes C# in memory via the Gemini API(fileless, detection-evading); APT42 used it to build phishing personas. Vendor defenses are "still holding for now," but malware calling an LLM at runtime is now an operational pattern. The Record· CyberScoop
Google I/O 2026 continues to land— Gemini 3.5 Flash(frontier performance for agents/coding, strong on long-horizon tasks), Gemini Omni(any-to-any: image/audio/video/text), and the Gemini Sparkagentic assistant (initially Ultra-exclusive); the Gemini App passed 900 million MAU (2x year over year), and Search's AI Mode passed 1 billion MAU. HeyGoTrade recap· Tom's Guide
Infrastructure— an 8th-generation TPU fork strategy: 8t(training; 9,600 chips per superpod, 2PB of shared HBM, 3x the compute of the prior generation) and 8i(inference; 1,152 chips per pod, 3x on-chip SRAM). Google Cloud Blog

This file is the standalone AI Frontier archive; the main brief is at intel/daily/2026-06-03.{zh,en}.md