Google Antigravity Review

Vendor risk isn't the question here. Google isn't disappearing in three years, full stop. That's rare in this category and it's worth something on its own.

The product itself is early. Browser-in-the-loop screenshots and the Antigravity 2.0 command center are genuinely differentiated versus Cursor or Copilot Workspace, but $20/month Pro tier and credit-based Ultra pricing at $100-$200 tells me they're still finding pricing-market fit. A public changelog exists, but no visible roadmap cadence that I could find.

Two questions before I'd scale this: is anyone on our eng team already fluent in Gemini-based tooling, and does browser-verification actually cut review time or just add another artifact to read. Pilot with a five-person frontend team for 60 days. Don't let it replace Copilot org-wide yet.

Artifact-based grounding — plans, walkthroughs, screenshots, browser recordings — is the right architecture for agent trust. If your team adopts this as the default review layer, in 3 years you've built review habits around Google's artifact schema, not an open one. That's a real dependency, not a hypothetical.

The permission/sandbox controls and allowlist/denylist model matter more than the demo screenshots. Compare that to Claude Code or Cursor's more editor-centric trust model — Antigravity is explicitly designed for background, unsupervised execution at scale, which is the harder engineering problem.

Pricing is credit-based with a $200/month Ultra 20x tier and enterprise gated behind a sales form — that's a lock-in signal, not a self-serve one. The MCP support and Python SDK keep the integration surface open enough to avoid full vendor capture, but the model access (Gemini 3 Pro, Claude Sonnet 4.5) still routes through Google's quota system.

$0, $20, $100, $200/month. Four tiers, visible without a sales call. Rare in this category — procurement won't need to chase a rep for the base product.

But the real cost isn't the tier. It's credits and quotas. Pro gives higher rate limits, not unlimited ones. Ultra 20x claims 20x Pro quota for $200 — that's a bet on usage patterns, not a fixed number. 50-dev team on Pro: $20 × 50 × 12 = $12K/year. Add Ultra upgrades when quotas run out mid-sprint, and year 3 is unpredictable, not $12K.

Compare to GitHub Copilot Workspace or Claude Code — similar credit-based ambiguity across the category. Enterprise tier requires a request-access form. No published overage rate that I could find. That's the invoice risk, not the sticker.

Free-tier quota refreshes weekly. Not hourly, not every five hours — weekly. Burn it through a Tuesday refactor and you're waiting days, not making coffee. The five-hour refresh starts at Google AI Pro, $20/month, which makes the free tier a trial rather than a daily driver.

Antigravity CLI is the piece that fits an existing setup: terminal-first, shell commands, background subagents driven from the keyboard, no migration into a new editor. Agents hand back screenshots and browser recordings rather than a diff you have to take on faith, which is the right shape for reviewing UI work.

The changelog shows steady releases — custom agent controls and keyboard navigation on September 18. But it all runs on desktop or in a terminal, so there's no phone check on a job you left running, and overage rides on purchased AI credits instead of a published rate.

The pitch here is delegation, not autocomplete. You describe a task, an agent goes and does it, and comes back with a plan, a walkthrough, screenshots, maybe a browser recording so you can actually see the UI change happened instead of trusting a changelog. That's a smarter handoff than most agentic tools bother with. Cursor and Claude Code mostly show you a diff and hope.

The pricing ladder is honest at least — free tier with rate limits, $20 Pro, up to $200 Ultra 20x (already cut down from $249.99, which tells you they're still finding the right number). No credit card on the free tier is a nice touch for a first ten minutes that doesn't feel like homework.

But there's no mobile and no web app beyond the marketing site; this is a desktop-and-terminal tool. Fine for now. Just don't call it 'always with you.'

Three tells. One: the free-tier limits have been cut more than once since the November 2025 launch. Two: a $249.99 sticker discounted to $200 this early. Three: 'command center' branding for what's basically a background-job monitor — the kind of naming that shows up right before a rename.

The feature list is real, not vaporware. Browser-in-the-loop verification, artifact grounding, MCP support, a Python SDK — that's more surface area than most Cursor or Claude Code competitors ship at launch. Pricing is concrete too: free tier, $20 Pro, $200 Ultra 20x, already discounted from $249.99.

Exit portability is the real question. IDE lock-in plus SDK-level custom agents means real migration cost if Google shelves this like Google Domains or Stadia. Google's track record with dev tools is mixed — Cloud Platform stuck, Inbox didn't. Enterprise tier exists, which helps. Early, but not nothing.