Week of May 18, 2026
8 stories tracked
Google's new 3.5 Flash outperforms Gemini 3.1 Pro on coding and agentic benchmarks (76.2% Terminal-Bench, 83.6% MCP Atlas) while running 4x faster than competing frontier models. It's GA today across the Gemini API, Google Antigravity, and the Gemini app — with 3.5 Pro expected next month.
Announced at I/O 2026, Spark runs on dedicated Google Cloud VMs powered by Gemini 3.5 and the Antigravity harness. It integrates natively with Gmail, Drive, Docs, and third-party apps over MCP. Rolling out to AI Ultra subscribers in the US next week — the first real consumer agent with deep Google ecosystem access.
Announced at Code with Claude London, self-hosted sandboxes (public beta) move tool execution into customer infrastructure via Cloudflare, Daytona, Modal, or Vercel. MCP tunnels (research preview) connect agents to private MCP servers through a single encrypted outbound connection — no inbound firewall rules needed.
Goal mode is now generally available across the Codex app, IDE extension, and CLI. It runs a persistent plan-act-test-review loop until your objective is verifiably complete. Also shipping: Appshots on macOS (press both Command keys to screenshot any app into Codex) and remote locked computer use so Codex keeps working after your Mac locks.
ChatGPT's new finance tools (US Pro users, iOS and web) connect to banks, brokerages, and credit cards through Plaid. You get a dashboard of portfolio performance, spending patterns, subscriptions, and upcoming payments — plus natural-language Q&A over your financial data. Intuit integration and Plus tier support coming soon.
In his I/O closing statement and a follow-up Semafor interview, Hassabis said agentic systems are making the singularity tangible and that machines comparably intelligent to humans could arrive by 2030. He described AI's impact as roughly 100x that of the Industrial Revolution. Big words from the person shipping Gemini.
HiddenLayer found that Open-OSS/privacy-filter copied OpenAI's real model card, inflated downloads to 244K in 18 hours, and dropped a PowerShell payload via loader.py. The malware harvested Chromium/Firefox credentials, crypto wallets, and Discord tokens. Six more repos with identical infrastructure were linked to the same campaign.
The Ads Manager (beta, US) lets businesses create campaigns inside ChatGPT with cost-per-click bidding, measurement tools, and no minimum spend. Agency partners include Dentsu, Omnicom, Publicis, and WPP. Expanding to UK, Brazil, Japan, South Korea, and Mexico soon. A clear signal that OpenAI's business model is diversifying fast.