AI Product Updates Daily — May 21, 2026

Today's digest covers the continued fallout from Google I/O 2026, a landmark hire at Anthropic, GitHub's double-barreled developer news, and an IPO filing that will reshape the AI sector's capital story. Three themes run through May 21: Google's subscription overhaul reaching consumers, agentic coding tools claiming the desktop, and enterprise AI deals getting bigger and faster.

Google

Gemini 3.5 Flash is generally available

Gemini 3.5 Flash launched into GA at Google I/O on May 19–20, now available through Google AI Studio, the Gemini API, and Android Studio.1 On the Terminal-Bench 2.1 coding benchmark it scores 76.2%, and on GDPval-AA it posts 1,656 Elo — both above Gemini 3.1 Pro — while costing under half the price of comparable frontier models. The model is optimized for long-horizon agentic tasks and is the default backbone for Google Antigravity.

Gemini 3.5 Pro is running internally and is expected to ship publicly next month.1

Gemini Omni is live for subscribers

Gemini Omni, Google's any-input-to-video model, is now available to all Google AI Plus, Pro, and Ultra subscribers via the Gemini app and Google Flow.1 Omni Flash is also free for YouTube Shorts Remix and the YouTube Create app for users 18 and older. The model generates video from text, image, audio, or video prompts and embeds an undetectable SynthID watermark in all outputs; verification is now live in Google Search and rolling out to Chrome in the coming weeks.

Google's subscription tiers are restructured and cheaper

Google overhauled its AI subscription lineup at I/O, with prices lowered across the board:2

Tier	Price	Key inclusions
Google AI Plus	$7.99/mo	Gemini 3.5 Flash, Gemini app with Omni, Google Flow with Omni, Daily Brief, AI Inbox in Gmail
Google AI Pro	$19.99/mo	All Plus features + YouTube Premium Lite ($8.99 value), Google Pics preview
Google AI Ultra	$99.99/mo (was $250/mo → was $200/mo)	All Pro features + 5× usage limits, 20 TB storage, priority Antigravity access, Gemini Spark (Beta, US only next week)

Quota now measures actual compute consumed rather than prompt count, and refreshes every 5 hours instead of daily. When the top-tier model cap is hit, the system falls back to a lighter model rather than cutting access.

Gemini Spark enters beta for Ultra subscribers

Gemini Spark, the 24/7 autonomous personal agent built on Gemini 3.5 and Google Antigravity, opens to US Google AI Ultra subscribers next week.3 It runs in the background after a device is locked, handles email and short-message replies, and can spawn custom sub-agents. Broader rollout continues through summer.

Daily Brief is live for US subscribers

Daily Brief, the overnight agenda-synthesis agent, launched today for all US Google AI Plus, Pro, and Ultra users in the Gemini app.3 It reads Gmail, Calendar, and tasks overnight, ranks priorities, and surfaces suggested actions each morning. A Google account connection is required.

Google Antigravity 2.0 ships with new developer tools

Antigravity 2.0 arrives as a standalone desktop app, bringing multi-agent orchestration: users can run a coding agent and a brand-asset generation agent simultaneously.1 Alongside it, Google released:

Antigravity CLI — terminal-native, no GUI required
Antigravity SDK — programmatic access to Google's agent framework, deployable on custom infrastructure
Managed Agents via Gemini API — a single API call spins up a remote Linux sandbox where an agent can browse the web, run code, and manage files; built on Gemini 3.5 Flash

Antigravity connects directly to Google Cloud projects under enterprise terms. Existing Gemini CLI users are advised to migrate; a migration guide is published.

Google AI Studio gets Android app development and Workspace integration

Starting today, Google AI Studio supports native Android app development directly from a build tab — including in-browser emulator preview, ADB debugging, and one-click publish to a Google Play internal test track.1 New developers get two free Cloud deployments with no credit card needed. The tool also now integrates with Google Workspace: apps built in AI Studio can read Drive files, Sheets data, and team Docs directly.

Gemini for Science opens in Google Labs

Three experimental research tools are available at labs.google/science starting May 19:1

Hypothesis Generation — multi-agent idea tournament for research questions (citations verifiable inline)
Computational Discovery — parallel evaluation of thousands of code variants for modeling tasks (epidemiology, solar forecasting)
Literature Insights — searches scientific literature, organizes results into a structured table, and exports to reports, slides, or audio summaries (built with NotebookLM)

Separately, Science Skills integrates 30+ life sciences databases (UniProt, AlphaFold, AlphaGenome API) as tools callable from Antigravity, compressing multi-hour bioinformatics workflows.

OpenAI

Codex mobile access is live inside ChatGPT

OpenAI's Codex coding agent is now accessible on iOS and Android through the existing ChatGPT app — there is no separate download.4 The mobile interface is a remote control for a desktop Codex session, not a mobile IDE: users can start tasks, approve or reject agent actions, and check progress from their phone. Files and credentials remain on the local machine; only diffs and status updates pass over a secure relay.

The pairing works with macOS Codex today; Windows support is listed as coming soon. All ChatGPT plans, including Free and Go, get access during this preview period — paid plans receive higher request limits.

Compared with Claude Code's mobile client (also inside the Claude app), Codex's differentiator is Computer Use: the desktop agent can interact with web apps and GUI tools beyond the terminal, while Claude Code counters with Agent Teams for multi-agent coordination on large codebases.

OpenAI Codex mobile — desktop pairing via QR code, showing synced task view on phone and VS Code session on laptop

OpenAI is preparing a confidential IPO filing

OpenAI is working with Goldman Sachs and Morgan Stanley to prepare a confidential IPO filing, targeting submission in the coming weeks.5 If filed shortly, the public offering could land as early as September. OpenAI's current private-market valuation stands at $730 billion from its most recent funding round. The company declined to comment beyond noting it "regularly evaluates a range of strategic options."

Anthropic

Andrej Karpathy joins to lead Claude pre-training

OpenAI co-founder and former Tesla AI director Andrej Karpathy confirmed on X that he has joined Anthropic, starting this week.6 He will report to Nicholas Joseph and lead a new team focused on accelerating Claude's pre-training using Claude itself.

"I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D. I remain deeply passionate about education and plan to resume my work on it in time." — Andrej Karpathy

The hire follows Anthropic passing OpenAI in business adoption (34.4% of companies tracked by Ramp's May 2026 AI Index, up from a lower base 15 months ago) and comes one week before Anthropic's Code with Claude developer conference.

Bristol Myers Squibb deploys Claude across 30,000 employees

Bristol Myers Squibb announced a strategic agreement on May 20 to position Claude Enterprise as its company-wide AI platform, rolling it out to more than 30,000 employees across drug discovery, clinical development, regulatory, manufacturing, and commercial operations.7 The deal makes BMS one of the largest pharma deployments of a frontier AI model to date.

GitHub / Microsoft

GitHub Copilot standalone app enters technical preview

GitHub launched the Copilot app in technical preview on Windows, macOS, and Linux for paid Copilot subscribers.8 The app operates entirely outside the IDE. Work starts from issues and pull requests already in a repository; each session has its own branch, file state, and conversation. Key capabilities include:

Agent Merge — monitors CI checks, downloads failure logs, resolves merge conflicts, and auto-pushes fixes before merge
Multi-session support — run multiple agent sessions across repos simultaneously, each isolated and trackable in real time
MCP server extensions — agents can be extended with custom skills and MCP servers

The move puts GitHub on the same desktop-client tier as Cursor and Claude Code.

GitHub Copilot billing switches to AI Credits on June 1

Starting June 1, GitHub Copilot will bill by AI credits — actual compute consumed — rather than a flat seat fee.9 Teams using premium models like Claude Opus or GPT-5 Codex at scale face bills roughly 9× higher than current flat-rate pricing under the previous plan. The change affects paid Copilot Individual and Enterprise seats; light users on simpler tasks will see smaller increases.

GitHub confirms 3,800 internal repos stolen through poisoned VS Code extension

GitHub disclosed on May 20 that an attacker gained access to approximately 3,800 internal repositories via a poisoned VS Code extension installed on an employee device.10 The stolen repositories contained infrastructure configurations, deployment scripts, and staging credentials. GitHub has detected and contained the breach, removed the malicious extension versions, isolated affected endpoints, and rotated compromised keys overnight.

The same threat group (TeamPCP / UNC6780 per Google Threat Intelligence) simultaneously hit the Nx Console VS Code extension (2.2M installs, stealing Claude Code credentials among others) and published 639 malicious npm package versions, and compromised Microsoft's durabletask Python SDK on PyPI — all within 24 hours, all in the Microsoft developer supply chain.

Recommended immediate actions for affected teams: rotate any plaintext secrets or GitHub tokens in private repos, pin GitHub Actions to full commit SHAs, audit VS Code extension auto-update policies, and rotate GitHub-issued OAuth app secrets and OIDC trust relationships.

Upcoming to watch

Gemini 3.5 Pro public release: expected in June
Anthropic Code with Claude developer conference: next week
Gemini Spark broader rollout and MCP third-party integrations: throughout summer
GitHub Copilot credit billing goes live: June 1
OpenAI IPO confidential filing window: coming weeks; potential public offering as early as September