If you’re already on Windsurf, upgrade. Wave 9 ships Windsurf’s first proprietary model family — SWE-1-lite replaces Cascade Base for free users at no extra cost. If you’re on Cursor, the calculus is tighter: switch only if you need JetBrains or Neovim support, or if enterprise compliance (FedRAMP, HIPAA, ITAR) is a hard requirement. VS Code users with autocomplete-heavy workflows should stay put.

Who this is for

Existing Windsurf users weighing whether Wave 9 is worth the update, and Cursor users deciding whether the agentic gap has closed. If you’re not already using an AI coding assistant and want a primer, start there first. If you’re comparing both IDEs side-by-side, see our Windsurf vs Cursor breakdown.

What we tested

Sources: Windsurf’s official Wave 9 release post (May 15, 2025), independent 2026 DEV Community review across a mid-size TypeScript codebase, Windsurf’s own internal benchmarks, and community reports from Reddit and Trustpilot. Resource usage figures (70–90% CPU on large repos during repository-wide edits) are drawn from the Second Talent 2026 review. Results vary by hardware and codebase size.

Windsurf app: Wave 9 build (May 15, 2025). Comparison anchor: Cursor v0.50, shipped May 24, 2025.

Findings

SWE-1 and flow awareness

Wave 9’s headline is Windsurf shipping its own model family for the first time. Three models:

Model	Who gets it	Quota
SWE-1	Pro users	0 credits (promotional at launch)
SWE-1-lite	All users, including free	Unlimited
SWE-1-mini	All users	Unlimited — powers Windsurf Tab autocomplete

SWE-1-lite replaces Cascade Base. That’s the meaningful upgrade for free users: a Windsurf-trained model with no monthly credit cap, in place of a third-party base model.

The core concept is flow awareness. SWE-1 is trained not just on the current file but on the developer’s full session timeline: text edits, terminal outputs, browser preview states, clipboard contents, IDE search history. The design bet is that engineers spend significant time outside the editor — in tests, terminals, and browsers — so a model that sees that full loop should perform better on agentic tasks.

On Windsurf’s own internal evals (Conversational SWE Task and End-to-End SWE Task), SWE-1 outperforms DeepSeek v3, Qwen 3, and Claude Haiku. These are internal benchmarks, not SWE-Bench — worth knowing before citing them in an architecture decision. SWE-1.5 (the successor model) later scored 40.08% on SWE-Bench, which matched Claude Sonnet 3.5 at that time. The frontier has since moved considerably — Claude Opus 4.5 now sits at 80.9%.

Cascade on complex multi-file tasks

Where Wave 9 earns its reputation is agentic multi-file navigation. An independent 2026 review documented Cascade finding 11 relevant endpoints across 4 router files during a refactoring session without any manual context-feeding. The comparison to Wave 8 is meaningful: Wave 8 (May 6, 2025) was a tooling release — Custom Workflows, file-based Rules, Simultaneous Cascades, MCP cap lifted from 50 to 100. Wave 9 replaces the intelligence underneath those tools.

The limitation is recovery. When a multi-step Cascade sequence crashes — most commonly during long Turbo Mode runs combined with background indexing — there is no partial state. Full restart required. On a long refactor, that means replaying context from zero. The issue is real enough that Windsurf’s own status page logged SWE model outages in May 2025, shortly after Wave 9 launched.

Resource usage adds to the problem on large codebases: 70–90% CPU and memory during repository-wide edits and multi-file refactoring. Files over 300–500 lines strain it further.

Pricing trajectory

At Wave 9 launch (May 2025):

Free: 25 credits/month, SWE-1-lite + SWE-1-mini unlimited, SWE-1 unavailable
Pro: $15/month, 500 credits, SWE-1 at 0 credits (promotional)

The $5 gap against Cursor’s $20 Pro was a real value advantage. It lasted ten months.

By March 2026, Pro moved to $20/month — exact Cursor parity. The credit system was replaced by daily/weekly quotas. Two new tiers appeared: Max at $200/month (Devin Cloud integration) and Team at $40/seat/month.

Wave 9’s pricing advantage is gone. The model story still stands, but you’re no longer getting the same capability for less.

What still needs work

Autocomplete quality is the weakest point. Windsurf’s inline completion hits 53–60% usability. Cursor and GitHub Copilot land at 70–75%. SWE-1-mini (which powers Windsurf Tab) triggers inconsistently and lags. If autocomplete acceptance rate is your primary productivity metric, Wave 9 does not close this gap.

Cascade has no partial recovery. Covered above, but worth repeating as a hard constraint: a crashed long-run agent sequence means a full restart. No checkpoint, no resume, no partial diff to save.

Reliability complaints persist. Trustpilot reviews trend 1-star: login failures, credits consumed on errored sessions, inconsistent output across identical prompts. This matches the documented May 2025 outages. It reads as scaling pain from a team growing fast, not a fundamental architecture problem — but it’s real friction in production use.

Benchmark transparency is thin. Wave 9 shipped with only internal evals. The external SWE-Bench number came later with SWE-1.5, and at that point the frontier had already moved past it.

IDE ecosystem is smaller. Windsurf supports 40+ IDEs — JetBrains, Vim, Neovim, Xcode. Cursor is a VS Code fork only. For non-VS Code users this is Windsurf’s strongest argument. For VS Code users, the plugin and keybinding ecosystem around Cursor is meaningfully more mature.

Verdict

Pick Windsurf Wave 9 if:

You’re on JetBrains, Neovim, or Xcode — Cursor is not an option for you
You work in regulated industries and need FedRAMP High, HIPAA, or ITAR compliance — Cursor offers SOC 2 only
Cascade’s multi-file agentic flow is more load-bearing than autocomplete speed
You were on the free tier and want SWE-1-lite at no extra cost

Stick with Cursor if:

You’re VS Code-native and autocomplete acceptance rate matters
You want async Background Agents (cloud VM, PR creation while you sleep) — Cursor v0.50 shipped these nine days after Wave 9
You rely on a rich ecosystem of VS Code plugins and community integrations
Raw market maturity matters: Cursor holds ~18% developer market share (JetBrains AI Pulse, January 2026)

Caveats

The crash-on-long-run behavior is hardware- and task-dependent — larger codebases and lower-spec machines hit it harder. The numbers in this review (53–60% autocomplete, 70–90% CPU) come from independent testing reported in mid-2025 and 2026 — not our own controlled benchmark.

Windsurf has a referral program: new Pro subscribers get 250 bonus flex credits via referral links, and referrers earn ~$10 in usage credit per successful referral. There is no revenue-share affiliate commission program. We have no affiliate relationship with Windsurf or Cursor; no links in this article earn us anything.

Windsurf rebranded from Codeium; windsurf.com now redirects to devin.ai following a subsequent acquisition. The Wave 9 SWE-1 architecture enabled later releases: Wave 13 (December 2025) added parallel multi-agent sessions, Wave 14 (January 2026) added Arena Mode and Megaplan.

References

Windsurf Wave 9 official announcement — primary source, May 15, 2025
Wave 8 Cascade customization blog — primary source, May 6, 2025
Windsurf pricing announcement — primary source
Windsurf models documentation — primary source
Windsurf status page — primary source, incident history
DEV Community 2026 review — independent review
Second Talent Windsurf review — independent review
UCStrategies benchmark roundup — benchmark comparison, SWE-Bench 40.08%
JetBrains AI Pulse survey January 2026 — developer market share data
SWE-1.5 official announcement — primary source, October 29, 2025
Windsurf Wave 13 announcement — primary source, December 24, 2025
Windsurf Wave 14 announcement — primary source, January 30, 2026
Claude Opus 4.5 announcement — Anthropic, SWE-bench Verified 80.9%