Three months ago, the AI IDE market was a clear two-player race: Cursor and Windsurf. Today it's three players with three different philosophies — and the brand "Windsurf" no longer exists. Here's the honest comparison after spending a week shipping production code in each.
The three players
| Vendor | Tool | Status |
|---|---|---|
| Anysphere | Cursor 3.0 | Shipped April 2, 2026; Cursor 3.5 cloud agents May 2026 |
| Antigravity 2.0 | Shipped May 19, 2026 at Google I/O | |
| Cognition AI | Devin Desktop (was Windsurf) | Rebranded June 2, 2026; Cascade retires July 1, 2026 |
Spec sheet
| Dimension | Cursor 3.0 | Antigravity 2.0 | Devin Desktop |
|---|---|---|---|
| Base | VS Code fork | VS Code fork | VS Code fork |
| Pricing (Pro) | $20/mo | $20/mo | $20/mo |
| Default model | Cursor-1 + Composer 2.5 | Gemini 3.5 Flash | Cognition SWE-1.6 |
| Multi-model | Claude, GPT, Gemini, Grok, DeepSeek | Google models only | Claude, GPT, in-house |
| Inline completion speed | Not published | 289 tok/s | 950 tok/s (Pro+) |
| Parallel agents | 8 sessions (Agents Window) | Yes (Manager view) | No |
| Worktree isolation | /worktree, /best-of-n | Yes | No |
| Cloud agent VMs | Yes (Cursor 3.5) | Yes (preview) | Yes (Devin cloud integration) |
| Browser testing agent | No | Yes (Browser Subagent) | No |
| Background scheduled tasks | No | Yes | No |
| JetBrains support | No | No | Limited (extension) |
| VS Code Marketplace | No (Open VSX) | No (Open VSX) | No (Open VSX) |
| MCP / Skills | Yes / Yes | Yes / Yes | Yes / Yes |
| Bring-your-own-key | Yes | Google-only | Yes |
| SOC 2 / Enterprise | Yes | In progress | Yes |
| Paying customers (reported) | 360,000+ | New | ~100,000+ |
Cursor 3.0 — the parallel-agent answer
Cursor 3.0 (April 2, 2026) reframed the IDE around the Agents Window.
You can open it with Cmd+Shift+P → Agents Window and run many
agents in parallel across:
- Local repos
- Git worktrees (one branch per agent, no merge conflicts)
- Cloud VMs (Cursor 3.5 adds isolated cloud execution, 8 parallel agents)
- Remote SSH
Two commands that matter:
/worktree— Fire an agent in an isolated git worktree. Changes don't touch your working tree until you accept./best-of-n— Run the same task in parallel across multiple models, each in its own worktree, then compare outcomes. This is the killer feature for refactoring where you don't know which model will produce the best result.
The Cursor 3.5 SDK (June 4, 2026) added:
- Custom stores (SQLite OR JSONL — readable + version-controllable)
- Custom tools (expose your own functions to the agent)
- Auto-review for local tool calls (classifier decides what needs approval)
- Nested subagents
Best for: Production work, multi-agent workflows, teams that want their own model routing across providers.
Weakness: No browser testing automation. No background scheduled tasks. The multi-agent UI is powerful but has a learning curve.
Antigravity 2.0 — the browser-aware IDE
Google's Antigravity 2.0 (May 19, 2026) is the most novel of the three. What makes it different:
Browser Subagent
A headless Chromium browser the agent drives while it builds. Write code → save → Subagent loads the running app → clicks through the UI → reports back what broke. This is the closest thing to "the agent is QAing itself" anyone has shipped.
For full-stack dev where you'd normally manually verify visual state, this is genuinely the first time the IDE replaces that step. For backend-only or pure-library work, it doesn't apply.
Manager view
Multi-agent orchestration similar to Cursor's Agents Window but designed around a "manager → worker" mental model. Tasks decompose into subtasks; you watch the manager assign them.
Background scheduled tasks
The only IDE in this group that ships scheduled background work — agents that run on cron, monitor your repo, file PRs proactively.
The model lock-in
Antigravity runs Google models only — Gemini 3.5 Flash today, Gemini 3.5 Pro when it ships in June. No bring-your-own-key. No Claude or GPT. This is Antigravity's structural weakness vs Cursor.
The quota problem
Antigravity 2.0 launched with an undersized quota system. Pro users ($20/mo) hit lockouts within hours of heavy use during the first week, triggering a public PR mess. Google has since increased limits but the brand damage is real for now.
Best for: Full-stack web dev where visual QA matters, teams all-in on the Google AI stack, developers who like background scheduled agent work.
Weakness: Single-vendor model lock-in. Quota uncertainty. CLI migration mandatory by June 18 (Gemini CLI → Antigravity CLI).
Devin Desktop — the IDE-plus-agent vision
Cognition AI acquired Windsurf in December 2025 for ~$250M and on June 2, 2026 announced the rebrand: Windsurf is now Devin Desktop. The Cascade agent (Windsurf's original autonomous engine) retires July 1, 2026; everything migrates to Devin's underlying agent runtime.
What changes
Devin Desktop = Windsurf IDE + Devin cloud agent integration:
- IDE-side editing + completion (unchanged from Windsurf)
- "Hand off to Devin" button → spawns a cloud Devin agent on the current task → executes asynchronously → returns with a PR
- SWE-1.6 in-house model for inline completion (950 tok/s on the Pro+ tier — fastest in the market)
- Standard multi-model support for the agentic side: Claude, GPT, in-house
Pricing wrinkle
Pro went from $15 to $20 with the acquisition. A new Max tier at $200/mo unlocks higher Devin cloud quotas + the SWE-1.6 Pro+ tier.
The "is this still Windsurf" question
If you used Windsurf before the acquisition, the answer is partially. The IDE shell is the same. The Cascade autonomous agent is gone. The new "hand off to Devin" workflow is more discontinuous than Cascade was — agents run in cloud VMs, not in your IDE session, and you wait for the PR.
Best for: Teams that want IDE + serious cloud agent in one brand, Devin-curious developers, fast-completion-priority users who'll pay for the SWE-1.6 950 tok/s.
Weakness: Naming + product churn. Cascade users have to migrate workflows by July 1. Devin's reliability vs Cascade is still being established.
The model story for each IDE
If you care which models you can actually run:
| Model | Cursor 3.0 | Antigravity 2.0 | Devin Desktop |
|---|---|---|---|
| Claude Opus 4.7 / 4.8 | ✅ | ❌ | ✅ |
| Claude Sonnet 4.6 | ✅ | ❌ | ✅ |
| GPT-5.5 (and 5.6 day-0) | ✅ | ❌ | ✅ |
| Gemini 3.5 Flash | ✅ | ✅ (default) | ✅ |
| Gemini 3.5 Pro (June) | ✅ (day-0) | ✅ (priority) | ✅ (day-0) |
| DeepSeek V4 Pro | ✅ | ❌ | ✅ |
| Bring-your-own-key | ✅ | ❌ | ✅ |
The model availability gap is the single most important factor in this comparison. Cursor and Devin Desktop both run every frontier model. Antigravity runs only Google. That's a structural tradeoff, not a temporary state.
Pricing reality at scale
The $20/mo Pro tier is the same across all three, but what $20/mo actually buys differs:
- Cursor Pro $20/mo: ~500 fast requests on Claude/GPT, then rate-limited to slow tier. Bring-your-own-key removes the limit entirely but you pay your own provider bill.
- Antigravity Pro $20/mo: Gemini 3.5 Flash only. Quotas recently increased but still tighter than Cursor for heavy use.
- Devin Desktop Pro $20/mo: Inline completion unlimited. Devin cloud agent runs limited to ~20/day.
For heavy users, Cursor + bring-your-own-key + Anvat as the model gateway is the cost-floor configuration in this set — you bypass Cursor's Pro quota and pay 30% off list price on every model.
What we recommend (use-case driven)
| If you're... | Use |
|---|---|
| Building React/Next.js full-stack | Antigravity 2.0 for the Browser Subagent — actually unique |
| Doing serious multi-model refactoring | Cursor 3.0 with /best-of-n |
| All-in on the Anthropic ecosystem | Cursor 3.0 + bring-your-own-key |
| Want serious cloud-agent automation | Devin Desktop (Devin handoff is the strongest cloud agent today) |
| Heavy daily completion user, want speed | Devin Desktop Pro+ for the 950 tok/s SWE-1.6 |
| Cost-conscious + heavy use | Cursor + bring-your-own-key + Anvat gateway |
| Open-source / self-hosted requirement | Aider + Cursor with local-model BYOK |
Where this goes next
The catalysts in the next 8 weeks:
- June 2026: Gemini 3.5 Pro ships → Antigravity 2.0's ceiling rises significantly (Browser Subagent + Pro-tier model is competitive with anything)
- Mid-late June 2026: GPT-5.6 likely ships → Cursor and Devin Desktop both day-0; Antigravity won't have it
- June-July 2026: Anthropic Mythos GA or Oceanus → Cursor and Devin Desktop will integrate; Antigravity won't
- Apple WWDC 2026: Xcode AI may launch → potential disruption for iOS-specific dev workflows
- September 2026: Anysphere → IPO rumored → Cursor pricing model may evolve
The dominant pattern: the multi-model IDEs (Cursor + Devin Desktop) keep their model-availability edge as new frontier models ship, while Antigravity has to bet on Google's model lineup being competitive at every release.
Related coverage
Cursor + bring-your-own-key + Anvat = the cheap pro setup
Cursor handles the IDE. Anvat handles the model bill. 30% off Claude / OpenAI / Gemini / DeepSeek list price, on one OpenAI-compatible key. Works with all three IDEs today.
Try free → →