Google shipped Antigravity 2.0 at I/O 2026, replacing the Gemini CLI with a five-surface agent platform: desktop app, CLI, SDK, managed cloud agents, and an enterprise tier1. Anthropic’s Claude Code spent the same period adding subagents, dynamic workflows, auto mode, and a 1M-token context window in its Q1 2026 feature wave7, then debugging three concurrent regressions that degraded output quality for seven weeks8. The two tools now compete for the same developers but represent divergent bets about how an AI coding agent should work.
Antigravity bets on the IDE. Its agent orchestrator decomposes tasks at runtime, spawns parallel subagents, and coordinates results from inside a desktop application1. Claude Code bets on the terminal. Its three-phase loop (gather context, execute, verify) runs sequentially from the command line, with subagents delegated into isolated context windows5.
The difference determines what each tool is good at, what it costs, and what breaks when you push it. This comparison covers the axes that affect daily work: pricing, model quality, agent orchestration, workflow automation, and the failure modes both vendors would rather not discuss.
The Architecture Bet: Terminal-First vs IDE-First
Claude Code is an npm package (@anthropic-ai/claude-code) that runs in any terminal. It reads files, searches codebases with Glob and Grep, talks to LSP servers for code intelligence, edits files, runs shell commands, and loops until the task is done5. The CLI is the primary surface. VS Code and JetBrains extensions exist but are thin wrappers around the same agent loop.
Antigravity 2.0 is the inverse. The desktop application is the primary surface; the CLI (agy, written in Go) replaces the now-sunset Gemini CLI4 and shares the same agent harness as the desktop app. An SDK exposes the harness programmatically in Python, TypeScript, and Go1. Unlike Antigravity 1.0, which was a VS Code fork, 2.0 is a standalone application with the IDE surfaced as a secondary view behind an “Open IDE” button added in a May 23 patch after launch-week backlash15.
The architectural difference shows up in practice. Hand Claude Code a request like “refactor the auth module and update the tests,” and it runs a single agent session: read the relevant files, plan the changes, edit them, run the tests, read the failures, iterate. Subagents can be spawned for isolated subtasks, but the orchestrator is always the main session and execution is sequential6.
Antigravity’s orchestrator decomposes the same request at runtime: one subagent handles the auth module refactoring, another updates the unit tests, a third verifies integration tests. These run in parallel with isolated working memory, and the orchestrator aggregates results and runs verification115. This is faster for tasks that decompose cleanly. It is worse for tasks where module A’s refactoring affects module B’s test assumptions, because the subagents do not share context by default16.
Pricing: $20/month Against Free, With Different Metering
Claude Code requires a Claude Pro subscription at minimum: $20/month ($17/month annual), which provides the CLI and Sonnet 5 as the default model. Max at $100/month multiplies capacity by 5x and adds 1M-token context windows on supported models. A Max 20x tier at $200/month offers 20x capacity. All plans use a 5-hour rolling window with weekly caps9.
Antigravity’s entry cost is zero. The free tier provides compute-based access to Gemini 3.5 Flash with usage limits that refresh every 5 hours until a weekly cap3. Google’s new billing model factors in prompt complexity, feature usage, and session length rather than counting raw requests3. Google AI Pro at $19.99/month raises the cap, and AI Ultra at $100/month offers 5x the Pro limit3.
The price points are structured to invite comparison. Claude Max at $100/month lines up against Google AI Ultra at $100/month. Claude Pro at $20/month nearly matches Google AI Pro at $19.99/month. Whether the underlying compute units deliver equivalent value depends on the model and how the vendor meters usage, which is not directly comparable.
The practical difference: Antigravity’s free tier is a usable productivity tool, not a trial mode. Claude Code’s free tier is too constrained for real work, so the effective floor is $20/month. For a five-person team, Claude Code Pro costs $1,200/year; Antigravity costs nothing until compute caps force an upgrade.
Model Quality: Claude Opus 4.8 vs Gemini 3.5 Flash
The model powering each tool is the largest determinant of output quality. Claude Code defaults to Sonnet 5 as of July 2026, with Opus 4.8 available on Max and above9. Antigravity defaults to Gemini 3.5 Flash, with Gemini 3.1 Pro available on higher tiers2.
On SWE-bench Verified, Claude Opus 4.8 scores 88.6%12. Gemini 3 Flash scores 75.80% on the official leaderboard12. The gap is real but its practical significance depends on the task distribution: the difference between 88% and 76% matters for complex multi-file bug fixes, but both models resolve the majority of straightforward issues. Terminal-Bench 2.1 provides a separate signal for shell-task performance13.
On the harder SWE-bench Pro dataset, Gemini 3.5 Flash scores 55.1% per Google’s model card2, while Claude Opus 4.7 scores 64.3%14. These are model-level scores. Neither Google nor Anthropic publishes product-level benchmarks that measure how the agent harness (tool orchestration, context management, error recovery) affects real-world performance relative to raw model intelligence.
On Terminal-Bench 2.1, Gemini 3.5 Flash scores 76.2% according to Google’s model card2. Claude Opus 4.8 scores in the mid-70s on the same benchmark variant per third-party evaluations. These are model-level numbers, not product-level agent scores, and neither vendor publishes how the agent harness affects real-world task completion. Terminal-Bench measures shell-centric tasks: compiling code, setting up servers, navigating file systems. It is a useful signal for CLI agents but does not capture the multi-file refactoring workflows where Claude Code’s sequential approach has an edge.
Google claims Gemini 3.5 Flash generates output tokens 4x faster than competing frontier models2. This is a vendor-reported figure with no independent replication. The speed advantage is plausible given Flash’s smaller parameter footprint, but the exact multiplier should be treated as approximate.
Anthropic suspended Claude Fable 5 and Mythos 5 on June 12, 2026, following a U.S. Commerce Department export control directive over national security concerns12. (Our coverage of the export order has the details.) Before the suspension, Fable 5 scored 95.0% on SWE-bench Verified. Anthropic’s most capable currently available model for coding is Opus 4.8.
Parallel Agents: Orchestration Models Compared
Claude Code supports four levels of parallelism: basic subagents (stable), Agent View for monitoring background sessions (research preview), Agent Teams with shared task lists (experimental, disabled by default), and Dynamic Workflows that script large subagent fleets with cross-validation (stable)6. Worktree isolation gives each parallel session its own git checkout, preventing file conflicts6.
Antigravity’s orchestrator decomposes tasks dynamically at runtime and dispatches to specialized subagents that run in parallel with isolated context1. Multiple agents can run simultaneously. This is an architectural advantage for tasks with clear decomposition boundaries, like generating frontend, backend, and test code in parallel.
The coordination cost is real. When two Antigravity subagents modify the same file, the last write wins silently with no merge conflict resolution16. Claude Code’s worktree isolation avoids this: parallel sessions operate on separate checkouts and merge through git. Antigravity’s documentation recommends scoping parallel tasks to non-overlapping file sets, which shifts the coordination burden onto the developer.
Claude Code’s subagents are defined by developers in markdown files (.claude/agents/*.md) with explicit system prompts, tool allowlists, and model selection6. This is more work upfront but produces predictable, auditable behavior. Antigravity’s dynamic decomposition requires less setup but offers less visibility into how the orchestrator split the task.
What Antigravity Does That Claude Code Cannot
Browser agent. Antigravity includes a managed Chrome instance the agent can control: launch dev servers, navigate pages, fill forms, take screenshots, and record video of the session1116. Walkthrough artifacts include browser recordings for visual review. Claude Code has no built-in browser testing; developers must shell out to external tools like Playwright.
Scheduled tasks. Antigravity supports cron-style scheduling via the /schedule command or a Scheduled Tasks panel11. Agents run in the background after the desktop app is closed and surface results as artifacts on next open. Claude Code has no built-in scheduling; teams rely on external cron or CI pipelines.
SDK for custom agents. The Antigravity SDK in Python, TypeScript, and Go exposes Tool.shell, Tool.code_edit, and Tool.web_search alongside custom system prompts and tool catalogs1. This lets teams build custom agent workflows on top of the Antigravity harness. Claude Code’s SDK exists but is narrower in scope.
Multi-folder projects. An Antigravity “Project” can include multiple folders (separate frontend and backend repos, for example) with isolated agent settings per project11. Claude Code works within a single repository root.
What Claude Code Does That Antigravity Cannot
Hooks. Claude Code’s hooks system runs user-defined shell commands, HTTP endpoints, or LLM prompts at fixed lifecycle points: session start, pre/post tool call, pre-response, post-compression10. These fire deterministically regardless of model behavior. A hook that runs eslint --fix after every file write always runs. Antigravity’s hooks in .agents/hooks.json cover fewer lifecycle events and are less documented.
Project memory. Claude Code reads CLAUDE.md files at session start, carrying project conventions, architectural rules, and coding standards across sessions without manual re-entry. Antigravity has an analogous GEMINI.md rules system, but third-party reviews report that agent context resets between sessions, losing iterative feedback captured in artifact comments16.
Auto mode. Claude Code’s auto mode routes tool calls through a model-based classifier that blocks irreversible operations while allowing safe actions to proceed without prompting7. This is a middle ground between manual approval and full autonomy. Antigravity offers four autonomy levels (Secure, Review-driven, Agent-driven, Custom) but the implementation is newer and less documented.
Plugin ecosystem. Anthropic maintains an official plugin marketplace and a community marketplace with reviewed submissions17. Plugins bundle skills (SKILL.md instruction injection), custom agent definitions, hooks, MCP servers, LSP servers, and background monitors into shareable packages. Antigravity has a skills system and MCP support but no equivalent marketplace.
Permission granularity. Claude Code asks permission before running destructive shell commands, writing to sensitive paths, or executing network requests. The sandbox applies only to the Bash tool, but permission prompts cover a wider surface. Antigravity’s four autonomy levels are configurable, but developers report frequent approval prompts in the default “Review-driven” mode that disrupt flow16.
Asterisks: Self-Reported Numbers and Known Failure Modes
Anthropic published a postmortem on April 23, 2026, admitting three bugs that degraded Claude Code output for seven weeks: a reasoning-effort downgrade from high to medium (March 4 to April 7), a cache-clearing fault that caused amnesia between turns (March 26 to April 10), and a system prompt change that constrained output and reduced coding quality by 3% (April 16 to April 20)8. Anthropic denied intentional degradation and reset usage limits for all subscribers. The seven-week gap between the first reports and the acknowledgment damaged developer trust.
Antigravity 2.0’s launch on May 19 had its own problems. The initial release shipped without a prominent IDE path. On May 22, Google reset weekly Gemini usage limits across all paid accounts, an implicit acknowledgment of capacity issues15. A May 23 patch added the “Open IDE” button and migration tool. Community reports cite login failures, agent hangs during peak hours, and browser-agent flakiness on complex single-page applications16.
Neither vendor publishes product-level agent benchmarks. Google reports model-level scores for Gemini 3.5 Flash. Anthropic reports model-level scores for Opus 4.8. How the agent harness affects real-world task completion relative to raw model scores remains unmeasured by either vendor. This is the gap where marketing meets engineering: a model that scores 88% on SWE-bench Verified can still produce an agent that fails 40% of real-world tasks if the orchestration layer is weak14.
Some third-party reviews claim Antigravity’s free tier offers access to non-Google models including Claude variants. This is not confirmed on any official Google pricing page or documentation. Treat it as unverified until Google publishes it.
How to Choose
Three questions decide the pick.
What model do you need? For complex multi-file refactoring, large codebase navigation, and code that compiles on the first try, Claude Opus 4.8 has a measurable lead on SWE-bench Verified12. For batch edits, test generation, and documentation, Gemini 3.5 Flash is faster and cheaper per interaction.
Where does your agent need to live? Terminals, SSH sessions, CI pipelines, and headless environments are Claude Code’s native habitat. A visual IDE with browser testing, scheduled tasks, and a multi-agent task manager is Antigravity’s strength. Some developers run both: Antigravity’s agy CLI as a subagent inside Claude Code is a configuration HN users have documented.
How much can you spend? Antigravity’s free tier has no equivalent on the Anthropic side. Claude Code’s effective floor is $20/month. At $100/month both Claude Max and Google AI Ultra offer premium capacity, but the compute units are not interchangeable.
For a solo developer evaluating AI coding agents, the pragmatic move is to start with Antigravity’s free tier and add Claude Code Pro when the compute limits or Gemini’s coding quality become a bottleneck. Teams with compliance requirements (audit logging, custom permission models, deterministic hooks) should evaluate Claude Code’s enterprise tier first.
Frequently Asked Questions
Can I use Antigravity and Claude Code together?
Yes. Developers have documented routing Antigravity’s agy CLI as a headless subagent from within Claude Code. This combines Antigravity’s parallel execution with Claude Code’s terminal-first orchestration, though it requires manual configuration and cross-vendor API coordination. The setup is not supported by either vendor.
Does Antigravity actually offer Claude models?
Some third-party reviews claim Antigravity’s free tier includes access to Anthropic’s Claude models. This is not confirmed on any official Google pricing page or documentation page as of July 2026. Treat the claim as unverified.
Which tool handles large codebases better?
Claude Code’s sequential, context-accumulating approach handles large monorepos more consistently. Its 1M-token context window on Max plans7 and LSP integration maintain structural understanding across sessions. Antigravity supports multi-folder projects but its agents start fresh each session, losing cross-session context unless manually captured in rules files.
What happened to Gemini CLI?
Google sunset Gemini CLI on June 18, 2026. The replacement is Antigravity CLI (agy), a separate Go binary. An antigravity migrate --from-gemini-cli command preserves skills, hooks, subagent definitions, and extensions4. Our earlier coverage of the sunset documented the migration timeline and pain points.
How do these compare to Cursor or GitHub Copilot?
Cursor is an IDE-native coding agent with its own model routing. GitHub Copilot is an IDE plugin that moved to token-metered billing in June 2026. Claude Code and Antigravity are agent platforms that operate outside any specific IDE. We covered the wider landscape and the April billing reshuffle separately.
