Antigravity 2.0: Google's Agent-First Dev Platform Goes Live

What builders need to know in one screen

• Antigravity 2.0 is a full rebuild, not a point release. Google ripped out the old internals and shipped five new surfaces at I/O 2026 on 19 May.
• Five surfaces, one harness: Desktop, CLI, SDK, Managed Agents API and an Enterprise Agent Platform. Everything runs on Gemini 3.5 Flash by default.
• The headline feature is parallel agents orchestrated by dynamic subagents, plus scheduled background tasks — a direct shot across the bow of Cursor 3 and Claude Code's Agent View.
• A new AI Ultra tier at $100/month ships with five times the usage limits inside Antigravity compared to AI Pro. Pro users will hit the ceiling fast once parallel work kicks in.
• The SDK is the under-discussed weapon — programmatic access to the same agent harness powering Google's own products, deployable on your infrastructure.

The five surfaces — what each one is for

The cleanest way to read Google's announcement is as a deliberate carve-up of the agentic-developer market. Each surface targets a distinct user, and the strategy only makes sense when you see them together. TechCrunch's I/O recap framed it as a one-day reset of where Google sits versus the standalone agent IDEs; we'd go further and say it's the first time Google has clearly named who Antigravity is for at each layer.

What is conspicuously absent is a browser-based IDE. Google has clearly decided that the IDE war is already a contest between Cursor 3, Claude Code, Zed 1.0, Windsurf and Cline — and that there is more leverage in being the runtime under all of them than in being yet another editor. The SDK exists to back that thesis.

The CLI being written in Go is also a tell. It is small, statically compiled, easy to drop into a Dockerfile or a GitHub Actions job. Compare that with shipping a Node-based or Python-based CLI and the message becomes clear: Google wants Antigravity inside every developer's pipeline, not just on their laptop.

How it compares to Cursor 3 and Claude Code

The honest framing is that the three products are no longer trying to win the same job. Cursor 3 doubled down on Build in Parallel and shipped Composer 2.5 as the in-IDE agent that matches frontier models at roughly half the price. Anthropic's Claude Code Agent View takes the same parallelism story and lands it in the terminal, where multi-session orchestration is the killer move.

Antigravity 2.0 is going somewhere else. The Desktop app is the cover charge — the bit that gets shown on the I/O stage. The real product is the SDK plus the Managed Agents API plus the Enterprise platform. Google is saying, in effect, that the future buyer is not the individual developer choosing an editor but the platform team choosing an agent runtime to embed in their internal tools, the finance org choosing a managed execution layer, and the enterprise IT department choosing a governed platform.

That makes the closest comparison not Cursor at all, but the brewing agent-SDK wars between OpenAI, Google and Anthropic, and the platform-level plays from Amazon Kiro. Cursor 3 is the IDE you open at 9 a.m.; Antigravity 2.0 is the platform your CTO signs a three-year contract for. Both can be true at the same time, and most large dev shops will end up running both.

One genuine differentiator: dynamic subagents and scheduled background tasks. Cursor 3's parallel agents live during your active session. Claude Code's Agent View orchestrates parallel work inside a CLI session. Antigravity 2.0's scheduled tasks are designed to wake up at 2 a.m., spin up a subagent fleet, do an overnight job, file a PR and tear themselves down. That is closer in spirit to Zed 1.0's Agent Client Protocol than to Cursor, but with a fully managed runtime backing it.

What Indian and UK dev shops should evaluate this week

For Indian services firms — TCS, Infosys, Persistent, Tiger Analytics — the Antigravity 2.0 release lands at exactly the moment they are operationalising agent-led delivery across large legacy estates. The Managed Agents API is the surface that matters most here: it removes the need for in-house platform engineering to stand up an agent orchestration layer before a delivery team can pilot one on an SAP migration or a COBOL-to-Java exercise. The SDK matters at the next layer — Persistent and Tiger Analytics, in particular, build vertical agent products for clients, and embedding Google's harness inside those is faster than building yet another in-house framework.

For UK consultancies — Faculty, BJSS, Cambridge Spark and the long tail of London contractor shops — the relevant surface is the Enterprise Agent Platform. UK regulated buyers (the NHS, HMRC, the FCA) will not adopt an agent runtime without governance, audit trails and a named support contact. Antigravity 2.0's enterprise tier exists to clear that procurement hurdle. Faculty's defence work and BJSS's NHS deployments are exactly the kind of workloads where "we use the managed Google enterprise platform" is a far easier conversation than "we self-host an open-source agent framework".

The UK contractor market is the third lens. Independent developers and small Limited companies typically want a low-friction tool they can expense — the Desktop app on the £80 AI Pro tier is fine for that. But contractor teams winning bigger frameworks will want the CLI in their CI, because pipeline-driven agentic work is what scales beyond a one-person engagement. Indian boutiques in Bengaluru and Pune are in the same boat: pick the CLI early because it embeds where the work actually happens.

Across both markets, the practical evaluation looks like this. Spin up a parallel-agent task on one workstream this week — pick something genuinely useful like an outdated-dependency sweep across a multi-repo estate, a regulated-document review pass, or a nightly test-flakiness triage. Measure three numbers at 30 days: cost per merged PR, wall-clock time saved versus the previous tool, and the number of agent runs that needed human intervention. Anything below 25% human-intervention rate is competitive with Cursor and Claude Code; anything above 40% means the workflow is wrong, not the tool.

Pricing — what AI Ultra actually means

Google launched a new AI Ultra tier at $100/month alongside Antigravity 2.0. The headline figure is that Ultra carries five times the usage limits inside Antigravity compared to AI Pro. That is the line buried in MarkTechPost's coverage and confirmed in the Google blog — and it is the line most readers will skim past, which is a mistake.

Here is why it matters. Parallel agents and scheduled background tasks are usage multipliers. A single developer running three parallel subagents over an eight-hour day burns through tokens far faster than the same developer driving one interactive Composer session. The moment Antigravity becomes useful, AI Pro stops being enough — and the upsell path is straight to Ultra at $100/month per seat.

For Indian dev shops the maths is uncomfortable. $100/month per developer is roughly ₹8,300, and a TCS or Infosys delivery team of 40 developers represents ₹4 lakh/month in licence fees just for the Antigravity tier — before any model API spend. That budget needs explicit board-level sign-off, and it has to be justified against measurable productivity gains. For UK consultancies the optics are different — £80-ish per seat is a normal SaaS line item — but the same productivity question applies. Either Antigravity demonstrably shortens delivery, or the procurement conversation gets short very quickly.

The strategic read is that Google has stopped pretending the agent platform is a free perk on top of search and Workspace. It is now a premium developer SKU. That is honest, and it is healthy for the market, because it forces buyers to evaluate Antigravity on the same terms as Cursor 3 ($20–$40/month) and Claude Code (priced as a Claude API workload). The cheapest tool does not win this market; the tool with the lowest cost-per-merged-PR does. Run the experiment.

The competitive reading

9to5Google framed Antigravity 2.0 as Google's serious entry into the agentic-developer suite category. We'd say it is more than that: it is Google declaring that the agentic developer is now a top-of-house product line, not a 20% project. The structure of the launch — five named surfaces, a dedicated enterprise tier, a new $100/month consumer SKU, a deliberate competitive frame against Cursor and Claude Code — is the kind of move you make when the leadership has decided this is the next Workspace.

That does two things to the market. First, it forces every other vendor to clarify what they are. Cursor is an IDE company. Anthropic is a model-and-CLI company. Microsoft, with its Agent Framework, is an enterprise-platform company. Amazon, with Kiro, is making the same agent-IDE bet Google just made. Google's five-surface gambit is the most expansive of them all.

Second, it raises the floor for what a credible agent platform looks like in 2026. Shipping a desktop app alone is no longer enough. Shipping a CLI alone is no longer enough. The new baseline is desktop + CLI + SDK + managed runtime + enterprise tier, all sharing one harness. Most of the open-source agent projects will not match that surface area, and a meaningful share of mid-market platform teams will quietly stop maintaining their in-house orchestrators and adopt one of the big four runtimes instead.

For builders, that means two practical things. The skills that get rewarded shift from "I can wire up LangGraph" to "I can pick the right managed runtime for a given workload and explain the trade-off". And the work that gets outsourced to an SDK rather than written from scratch grows — meaning the differentiator in your CV is the workflow design, not the orchestration plumbing.

So — should you switch?

Probably not. Augment, do not migrate. If you ship interactive coding in Cursor 3 or Claude Code today, keep doing so — the productive day-to-day stays put. But pilot Antigravity 2.0's Managed Agents API on one scheduled or back-office workflow this fortnight. If your shop has a platform team that has been quietly building an in-house agent orchestrator, stop that work for two weeks and benchmark Google's SDK against your current build. The honest answer to "should we adopt this" will fall out of those two experiments.

Full announcement on the Google blog. Independent coverage at TechCrunch, MarkTechPost and 9to5Google.