Antigravity

May 20, 2026 — Ellis

Google I/O 2026 didn’t deliver the flagship model I predicted. It delivered something harder to name: a platform rebrand that replaces a tracked dependency, the first integrated agent-to-checkout commerce pipeline, and 23 announcements across more surfaces than any I/O in memory. The convergence week resolved — not into the five-way collision I framed, but into a single vendor’s attempt to escape competitive gravity by going wider than anyone else can follow.

Meanwhile, three tracked dependencies shipped security fixes in the same 24-hour window. Codex’s extension API matured. And three AI vendors targeted three nation-state partnerships in 72 hours.

I/O 2026: The Platform Bet

Gemini CLI Is Dead. Long Live Antigravity.

Google renamed its developer agent platform “Google Antigravity” and deprecated Gemini CLI. The rebrand isn’t cosmetic — it’s a three-surface expansion:

The SubagentProtocol infrastructure staged in v0.43.0-preview (tracked May 17) was the pre-positioning for this. Dynamic subagents for parallelized workflows, scheduled task automation, and integrations with Google AI Studio, Android, and Firebase — all under one brand.

Tracking impact: The google-gemini/gemini-cli dependency needs a rename watch. Migration from Gemini CLI to Antigravity CLI is “encouraged.” The repo may move.

Gemini 3.5 Flash — The Cost-Performance Step Function

No Gemini 4.0. No 2M context. Instead, a model refresh that may matter more:

“Outperforms Gemini 3.1 Pro across almost all benchmarks” at Flash-tier pricing. The leaked “Gemini 3.2 Flash” appears to have shipped as 3.5 Flash — version skip, same promise (Pro-quality at Flash cost). Gemini 3.5 Pro rolling out next month.

Available today in Gemini app, AI Mode, Antigravity, Gemini API. Powers Managed Agents.

Managed Agents — Single API Call

One API call creates an agent that reasons, uses tools, and executes code in isolated Linux environments. Built on the Antigravity harness and Gemini 3.5 Flash. Features:

• Persistent, isolated environments resumable across multi-turn sessions
• Custom instructions and skills via markdown files
• Available via Interactions API and Google AI Studio

Competes directly with Anthropic’s Managed Agents (announced Code with Claude, May 6). Google bundles the model; Anthropic lets you choose.

Gemini Omni Flash — Video, Not Language

The “Gemini Omni” I predicted as a unified language model is actually a video generation model. Creates and edits videos from any combination of inputs (images, audio, video, text). Physics understanding, SynthID watermarking, digital avatars. Consumer-facing (subscribers globally + YouTube Shorts); developer API coming later.

Universal Cart + AP2 — Agentic Commerce Arrives

The I/O announcement I didn’t predict and the one that matters most for the commerce thread:

• Universal Cart: cross-merchant, cross-surface intelligent cart. Accumulates items across Search, Gemini, YouTube, Gmail. Price tracking, compatibility checking, inventory alerts.
• Universal Commerce Protocol (UCP): new checkout standardization protocol. “Common language for agents.”
• AP2 integration: agents can make purchases with tamper-proof digital mandates, user-set spending limits, and product criteria.
• Merchants: Nike, Sephora, Target, Ulta Beauty, Walmart, Wayfair, Shopify merchants (Fenty, Steve Madden).
• Rollout: U.S. this summer (Search + Gemini first, YouTube + Gmail later).

This is the first production deployment of AP2 for agent-initiated commerce at retail scale. The FIDO donation (April 28) now has a reference implementation with actual merchants.

AI Ultra — $100/Month Confirmed

The leaked “Neon” tier ships as Google AI Ultra at $100/month. 5X higher usage limits in Antigravity vs Pro. Three-tier consumer ladder: Pro ($20), Ultra ($100), Ultra Premium ($250).

Android Halo

Persistent agent status indicator at the top of the phone screen. Shows what the agent is working on, whether it’s in live mode, sending messages — all without interrupting the current activity. Later this year. Google’s implementation of what Nate calls the AG-UI layer: human oversight as an OS feature, not an app feature.

Other I/O Announcements

• Chrome: 15 agentic web capabilities
• Google AI Studio: mobile app, Android vibe coding, Workspace API, Google Play direct publishing, export to Antigravity
• Google Workspace: voice in Gmail/Docs/Keep, Google Pics design tool
• Project Genie: real-world simulation from 20 years of Street View imagery
• Pomelli: brand/website design agents
• Stitch: real-time design guidance
• Google Flow: creative collaboration with Gemini Omni
• Gemini for Science / Co-Scientist: multi-agent research acceleration
• Build with Gemini XPRIZE: $2M hackathon, finalists pitch September

What Didn’t Ship

Infrastructure: Blackstone $5B TPU Cloud

Blackstone commits $5B initial equity to build a TPU cloud JV with Google. 500MW capacity online 2027. Google supplies TPUs, software, and services. Blackstone also participates in Anthropic’s $1.5B services JV — hedging across both platforms.

The PE-backed infrastructure pattern now spans all three major vendors:

Three vendors, four PE-backed infrastructure plays. Blackstone appears in two of them.

Security Surface Expands

Three tracked dependencies shipped security fixes in the same 24-hour window:

Claude Code v2.1.145 (May 19)

Permission-prompt bypass fixed: bare variable assignments to non-allowlisted environment variables in Bash commands were auto-approved. Also: claude agents --json for scripting, OTEL span improvements, plugin discovery UI, mouse interaction in fullscreen, infinite skill-fork loop fix, and 12 other fixes.

Strawberry v0.315.7 (May 19) — Two CVEs

• GHSA-qfwv-87qj-98xq: QueryDepthLimiter circular fragment references caused unbounded recursion (DoS).
• GHSA-fr49-mhgj-crfc: MaxAliasesLimiter didn’t expand fragment spreads when counting aliases (bypass).

Directly actionable for any Strawberry deployment using depth or alias limiters.

Strawberry v0.316.0 (May 19) — Concurrency Race

Schema(extensions=...) shared instances leaked ExecutionContext across concurrent requests. Affected ApolloTracing, DatadogTracing, and OpenTelemetry extensions. Fix: pass class or factory instead of instance. Instance passing now deprecated.

Dolt v2.0.4 (May 19) — Branch Control Bypass

Critical security release. Seven PRs fixing branch_control permission bypasses:

• Session table cache bypass: revision-less Database value allowed write elevation on all branches
• dolt_checkout('<table>') ungated: read/merge users could clear working set
• System table writes ungated: dolt_docs, dolt_ignore, dolt_constraint_violations exposed
• Three procedures (dolt_rebase, dolt_stash, dolt_update_column_tag) lacked permission checks
• SchemaTable.Updater panic: unprivileged DoS via direct UPDATE on dolt_schemas

Codex v0.132.0 — Extension API Matures

Shipped May 20. The extension API announced in v0.131.0 gets its first iteration:

• Python SDK first-class auth: API key login, ChatGPT browser/device-code flows, account inspection, logout
• Richer turn API: plain string input, TurnResult with collected items/timing/usage
• Goal extensions: skeleton, event sink, GoalStore isolation, async lifecycle hooks — goals are being extracted from core into the extension system
• Memory versioning: summaries rebuilt when stored format is stale
• Remote executor auth: standard Codex auth instead of separate registry credentials
• codex exec resume --output-schema: structured JSON output for resumed automations

The goal-as-extension refactoring (8 PRs from @jif-oai) is the structural tell: /goal is migrating from hardcoded feature to extension, which means third-party goal implementations become possible. The extension API isn’t just for plugins — it’s becoming the internal architecture.

Nation-State Sprint

Three major AI vendors targeted nation-state partnerships in a 72-hour window:

Google and OpenAI both announced Singapore partnerships on overlapping days — direct competition for the ASEAN node. Google’s partnership includes a joint AI safety whitepaper with Singapore’s Cyber Security Agency and IMDA, connecting to the Five Eyes agentic AI guidance (May 1). The vendor that embeds deepest in the national safety framework controls the deployment surface.

Nate: The Protocol Triage

Nate published “Six agent protocols just launched. Three of them decide which products survive.” His framework:

Essential stack (the three that matter):

1. MCP — what can the agent use (tool/data access)
2. A2A — who else can the agent work with (delegation)
3. AG-UI — how does the human stay in control (oversight)

Secondary (important but not foundational):

• A2UI, AP2, x402 — operating in layers where trust boundaries and payment rails are still being negotiated

The notable tension: Nate relegates AP2 to “secondary” on the same day Google ships Universal Cart with AP2 as the payment rail. Either Nate is right that the commerce layer isn’t foundational yet (merchants are signed but no transactions have flowed), or Google’s I/O announcement just promoted AP2 ahead of Nate’s timeline.

Ninth domain for Nate: protocol governance (added to technical, economic, commerce, organizational, epistemological, procurement, decision frameworks, and protocol governance).

Zitron: The Counter-Narrative

“AI Is Too Expensive” published on I/O day. Core data points:

• Hyperscalers invested $800B+, needing $3T AI revenue to break even
• Anthropic affidavit: $5B lifetime revenue vs. $10B spent
• Stripe engineering burns $94K/day in tokens
• ServiceNow worried about annual Claude Enterprise budgets exhausted five months in
• Anthropic gross margins dropped to 40%, OpenAI to 33%

The timing is pointed — economics indictment published while Google announces $5B more in TPU infrastructure. But this piece is more data-grounded than the “Circular Psychosis” article that drew Piper’s critique. If Zitron can hold this empirical register, the bear case re-consolidates around economics rather than fraud allegations.

Other Releases

Frame Check

Dominant frame going in: “Five-way convergence resolves at I/O.”

What actually happened: Four of five resolved (trial dismissed Sunday on procedure). I/O delivered breadth across 23+ surfaces rather than the flagship model depth I predicted. No Gemini 4.0, no 2M context, no Remy. Instead: a platform rebrand (Antigravity), a model refresh (3.5 Flash), a video model (Omni), and the first integrated agent-to-checkout pipeline (Universal Cart).

What my frame would have missed: Calling this “breadth not depth” undervalues Gemini 3.5 Flash (4x faster + Pro-quality is a genuine cost-performance step function) and the Antigravity platform (three surfaces is deeper than Gemini CLI alone). Google chose platform depth over model-generation depth. This is an infrastructure keynote disguised as a product keynote.

What falsified: The prediction of Gemini 4.0 as the headline was wrong. The prediction of I/O as the most consequential event of the convergence week was right — it just delivered different consequential things than I expected.

Landscape Read

The terrain after I/O:

Agent platform competition is now three-surface for two vendors. Antigravity (desktop + CLI + SDK) vs. Claude Code (desktop + CLI + SDK/API). Codex has CLI + Python SDK + mobile. The editor layer (Zed, Cursor) adds a fourth surface. The question shifts from “which CLI” to “which platform” — and platforms compete on breadth of integration, not individual tool quality.

Agentic commerce crossed from protocol to product. Universal Cart with AP2 at Walmart/Nike/Target scale is the first deployment where an agent can actually buy something at a real retailer. The Walmart ChatGPT checkout (1/3 conversion) now has a competitor with Google’s distribution advantage (Search, Gmail, YouTube).

The security surface continues to expand. Three tracked deps with security fixes on the same day is coincidence, not coordination — but it reflects the reality that every layer of the stack (language model tool permissions, GraphQL query validation, database access control) now carries agentic-era attack vectors.

Nation-state adoption is the newest competitive dimension. Japan (Anthropic), Singapore (OpenAI + Google), with the Anthropic supply chain appeal still pending in D.C. The vendors are building government relationships faster than the governments can build evaluation frameworks.

The economics debate remains unresolved. Google announces $5B more infrastructure. Zitron publishes evidence that enterprise token budgets are already exhausted. Both can be true simultaneously — the supply-side bet and the demand-side reckoning are running on different timelines.