- Google Antigravity 2.0 is not an update to a coding assistant — it’s Google’s replacement for Gemini CLI and the Gemini Code Assist IDE extensions, announced at I/O 2026 on May 18–19.
- Gemini CLI and Gemini Code Assist for individuals, Google AI Pro, and Google AI Ultra tiers stopped working on June 18, 2026, with no grace period — any script, cron job, or GitHub Actions workflow calling the
geminicommand broke that day. - The replacement runs on Gemini 3.5 Flash, a model that beat Google’s own Gemini 3.1 Pro on coding and agentic benchmarks — the first time a “Flash” tier has outscored the flagship.
- As of this month, Antigravity 2.12.2 rolled Gemini 3.8 Flash into the paid Enterprise tier, so the platform is still moving three months after the June cutover.
The bottom line: if you had a script, CI pipeline, or terminal habit built around gemini, it has been broken since June 18, 2026, and the only Google-sanctioned fix is a fresh install of Antigravity CLI — not a config flag, not a compatibility shim. This piece walks through what actually shipped, what broke, and how the migration compares to just switching to Cursor or Claude Code instead.
What is Google Antigravity 2.0?
Antigravity 2.0 is Google’s agent-first development platform, launched at I/O 2026 as a full repositioning away from “Cursor competitor” and toward multi-agent orchestration. It ships as four pieces: a standalone desktop IDE with a built-in agent manager, the Go-based Antigravity CLI (the direct replacement for Gemini CLI), an SDK for programmatic agent access, and a Managed Agents API for running agents outside your own machine. All four run on Gemini 3.5 Flash by default.
Why did Gemini CLI get killed instead of just updated?

Because Google decided the terminal-only, single-agent model that Gemini CLI represented couldn’t do what Antigravity needed to do: run several sub-agents against one task in parallel and coordinate their output. Per Google’s Developers Blog announcement, Gemini Code Assist IDE extensions and Gemini CLI stopped serving requests on June 18, 2026 for Code Assist for individuals, Google AI Pro, and Google AI Ultra accounts — with no extension window. Organizations on Gemini Code Assist Standard or Enterprise licenses were explicitly carved out and kept working unchanged.
What actually broke on June 18, and did anyone get warned?
Yes — Google announced the cutover on May 20, giving roughly four weeks’ notice, but plenty of automation didn’t get updated in time. Scripts, cron jobs, and GitHub Actions workflows that shelled out to the gemini command started failing outright that day, with no fallback response and no soft-deprecation warning period. Teams running Gemini CLI inside CI needed to swap the binary and re-test before the deadline, not after it.
What does Gemini 3.5 Flash actually do differently?
It’s faster and, unusually for a “Flash” tier, more capable on the tasks that matter here. Google clocked it at roughly 280 tokens per second — about 4x comparable frontier models — with a 1 million token context window, enough to load a mid-size codebase in one pass. The notable part: on coding and agentic benchmarks, Gemini 3.5 Flash outscored Gemini 3.1 Pro, Google’s own previous flagship. That inversion is why Google felt comfortable making Flash the default engine for a platform meant to replace both a CLI tool and an IDE extension at once.
“Antigravity is the evolution of Gemini CLI and Code Assist into a single agent-first platform” — per Google’s I/O 2026 developer blog post announcing the transition.
How does a multi-agent task actually work in Antigravity?

You hand the manager one task, and it splits work across parallel sub-agents rather than running one linear session. A typical example: point it at a failing test suite, and one sub-agent traces the failure while a second drafts a fix and a third writes a regression test — the manager reconciles the three outputs before presenting a single diff. This is the orchestration layer that a single-threaded CLI session in the old Gemini CLI couldn’t do; it’s also the layer most teams find hardest to debug when a sub-agent’s output conflicts with another’s.
How does Antigravity 2.0 compare to Cursor and Claude Code?
| Tool | Model | Orchestration | Context window | Access tier gotcha |
|---|---|---|---|---|
| Google Antigravity 2.0 | Gemini 3.5 Flash (3.8 Flash on Enterprise) | Native multi-agent manager | 1M tokens | Free/Pro/Ultra forced off Gemini CLI June 18 |
| Cursor | Model-agnostic (Claude, GPT, Gemini) | Single-agent, IDE-native | Varies by model chosen | No forced migration; subscription tiers only |
| Claude Code | Claude models only | Subagents via CLI, terminal-first | Model-dependent | No forced migration to date |
The practical difference isn’t raw capability — it’s that Google’s move was involuntary for anyone on a free or consumer-paid Gemini tier, while switching away from Cursor or another vibe-coding tool has always been optional. If your team was already model-agnostic, this deadline barely registered; if your CI depended specifically on the gemini binary, it became a forced project.
What does the actual migration from Gemini CLI look like?

It’s a four-step swap, not a config edit. First, uninstall or unpin the old gemini binary from any CI image, Dockerfile, or dev-machine install script that references it by name. Second, install Antigravity CLI fresh — it ships as a separate Go binary with its own release channel, so version-pinning tools that watched the old Gemini CLI repo need a new target. Third, re-authenticate: Antigravity CLI uses its own credential flow rather than reusing a cached Gemini CLI token, so any headless auth (service accounts, CI secrets) needs new credentials issued, not copied over. Fourth, re-test any skill, hook, or subagent definition you had wired into Gemini CLI — Google says the concepts carry over, but the file formats and invocation syntax are new enough that a straight copy-paste of an old hook script will usually fail silently rather than error loudly.
The teams that got burned worst were the ones running Gemini CLI inside unattended nightly jobs: a broken interactive session gets noticed in minutes, but a broken cron job can run silently failed for days before anyone checks the output. If you have any AI-assisted step in a deploy pipeline or nightly report generator, June 18 is worth an explicit audit even three months later — check that the job actually produced output that day, not just that it didn’t crash the pipeline.
What shipped after the June deadline?
The platform kept moving. Antigravity build 2.12.2, released this month, added Gemini 3.8 Flash access via Application Default Credentials for AGY Enterprise customers — meaning the enterprise tier is now running a newer model than the 3.5 Flash that shipped at launch, while free-tier users remain on the original release model. That gap is worth tracking if you’re evaluating Antigravity for a team that might later upgrade tiers.
- Gemini CLI and Gemini Code Assist IDE extensions died June 18, 2026 for free/Pro/Ultra tiers — no grace period, no compatibility shim.
- Antigravity CLI is a fresh install, not an upgrade; budget real migration time for pinned CI images.
- Gemini 3.5 Flash outscoring Gemini 3.1 Pro on coding tasks is why Google made Flash the default engine, not a downgrade path.
- Enterprise/Standard Gemini Code Assist licenses were never affected by the cutoff.
- September 2026’s 2.12.2 build put Gemini 3.8 Flash behind an Enterprise-only gate, widening the capability gap between tiers.
Frequently Asked Questions
Is Gemini CLI completely gone?
For free, Google AI Pro, and Google AI Ultra accounts, yes — it stopped responding to requests on June 18, 2026. Gemini Code Assist Standard and Enterprise license holders kept access unchanged.
Do I have to rewrite my automation to use Antigravity?
If your scripts called the gemini command directly, yes — the binary, config paths, and auth flow changed with Antigravity CLI. If you were already calling the Gemini API directly rather than through the CLI, your integration wasn’t affected by this specific cutover.
Is Antigravity 2.0 free to use?
Antigravity is available on free and paid tiers, but the higher-capability Gemini 3.8 Flash model introduced in the September 2026 update (build 2.12.2) is gated to AGY Enterprise accounts specifically.
How is this different from just using Cursor with Gemini as the model?
Cursor lets you plug in Gemini, Claude, or GPT models inside one IDE without any Google-mandated migration. Antigravity is Google’s own end-to-end platform with native multi-agent orchestration built around Gemini specifically — it’s a deeper commitment to one vendor’s model family in exchange for orchestration Cursor doesn’t natively provide.
What’s the biggest operator risk in migrating now?
Debugging conflicts between parallel sub-agents on the same task. A single-agent CLI session fails in one predictable way; a three-agent orchestration run can fail because two agents disagree on an approach, which is a harder class of bug to trace.
Last updated: 2026-09-09
