Is Cursor 3.7 or Cursor 4 the current version in June 2026?

Both are shipping at the same time and the answer depends on which release channel you're on. Cursor 3.7 with Composer 2.5 is the stable channel build that most users are on today — it's the version reviewed in the prommer.net and Cosmic JS comparisons published in mid-June 2026. Cursor 4 is the next major release that the team announced earlier with auto-router model selection and an updated SDK, and it has been rolling out gradually to power users on the early-access channel. As of June 16, 2026 most paying Cursor users are on Cursor 3.7. Cursor 4 GA is expected within weeks. If a tutorial or comparison page is dated mid-June 2026 and mentions just 'Cursor', it almost certainly means 3.7.

What's the actual difference between Cursor 3.7 Composer 2.5, Windsurf SWE-1.5, and Claude Code v2.1.169?

Three different design centers. Cursor 3.7 Composer 2.5 is an editor-first agentic mode — you drive the IDE, Composer handles multi-file edits and orchestrates models through Cursor's tab-completion-trained selector. Windsurf SWE-1.5 is Cascade-driven and emphasizes hands-off long-running tasks; the SWE-1.5 in-house model is the default. Claude Code v2.1.169 (June 8, 2026 release) is terminal-first; the killer features this month are the --safe-mode flag that disables custom skills and MCP servers for safety-sensitive sessions, plus background sub-agents that run autonomously. Picking one: editor-comfort developer → Cursor; hands-off-long-tasks → Windsurf; terminal-first autonomous workflow → Claude Code.

Which is fastest on SWE-bench Pro in June 2026?

Numbers as published this month: Claude Code with Claude Fable 5 leads on SWE-bench Pro at roughly 69.2% per Morph LLM's June 2026 ranking, with Codex CLI at the bench ceiling for Terminal-Bench specifically and Cursor and Windsurf trailing slightly behind. SWE-bench Pro is a proxy, not a guarantee — Cursor 3.7 with Composer 2.5 routing to Fable 5 closes most of the gap because the model is the same; the difference is the agent loop. For most practical work, the editor or the terminal you're comfortable in matters more than 5-point benchmark swings. For autonomous overnight sub-agent work, Claude Code is the right tool because the sub-agent surface is more mature.

What does Claude Code v2.1.169's --safe-mode flag actually do?

It launches Claude Code with all customization disabled — no custom Skills loaded, no MCP servers connected, no project-level CLAUDE.md instructions applied, and the agents directory ignored. The use case is debugging: when something behaves unexpectedly, --safe-mode tells you whether the issue is in Claude Code core or in your customizations. The env variable CLAUDECODESAFEMODE works equivalently. The same v2.1.169 release also disables GPU acceleration in VS Code/Cursor/Windsurf integrated terminals to fix garbled-text rendering issues, and adds substring matching to the claude agents slash-command autocomplete. It is a maintenance release, not a feature release, but the safe-mode flag is the one practical addition worth knowing if you live in Claude Code.

Quick Answer

Cursor 3.7 vs Cursor 4 vs Windsurf vs Claude Code: What's Live (June 2026)

Published: June 16, 2026

Cursor 3.7 vs Cursor 4 vs Windsurf vs Claude Code: What’s Live (June 2026)

Three of the four leading AI coding tools shipped meaningful updates in the past two weeks, and the version numbers being thrown around in comparison posts are inconsistent enough to be confusing. This page maps what is actually live in June 2026, what each just added, and which to use for what.

Last verified: June 16, 2026.

TL;DR

Cursor 3.7 with Composer 2.5 is the stable build most users are on right now.
Cursor 4 is rolling out on early-access; full GA is imminent but not yet here.
Windsurf SWE-1.5 is the current Windsurf model and Cascade is the agent loop.
Claude Code v2.1.169 (June 8, 2026) added —safe-mode plus terminal fixes.
Claude Fable 5 is the model behind every leading agent’s recent benchmark gain; access ends in the unified claude.ai plan on June 22, then moves to usage credits.

Version state of play right now

Tool	Current stable	What’s next	When
Cursor	3.7 + Composer 2.5	Cursor 4 (early-access live)	GA expected within weeks
Windsurf	SWE-1.5 model + Cascade	Continuing point updates	Rolling
Claude Code	v2.1.169	Continuing weekly releases	June 8 was last major
Codex CLI	Latest	Continuing	Rolling

The two confusing items are Cursor and Claude Code. Both ship aggressively — Cursor on rotation between 3.x and 4.x branches, Claude Code on a roughly weekly cadence — so any review more than two weeks old is already partly stale.

Cursor 3.7 with Composer 2.5

Composer 2.5 is Cursor’s flagship agentic mode and the headline feature of the 3.7 branch. It does multi-file edits across a project given a natural-language instruction, with an in-line plan that you can edit before execution. The tab-completion model is custom-trained by Cursor specifically for the editor.

What’s new vs Composer 2:

Faster planning step (down from ~8s to ~3s on cached projects).
Better long-context handling — Composer 2.5 can reason over an entire mid-sized repo without context-window thrashing.
Auto-router-light: picks between Claude Fable 5, GPT-5.5 Pro, and Cursor’s in-house tab model depending on the task.

Pricing: Cursor Pro at $20/month, Cursor Business at $40/seat/month. Composer usage counts against monthly inclusion; heavy users hit limits.

Best for: Developers who want an IDE-first workflow with strong tab completion and a Composer for medium-complexity multi-file work.

Cursor 4 (early-access)

Cursor 4 is the next major branch, currently rolling out to a subset of paying users via the early-access channel. Confirmed features:

Full auto-router that picks between models on a per-prompt basis without needing manual selection.
Updated Cursor SDK for building agents on top of Cursor primitives.
Cloud Composer — a hosted Composer surface for long-running tasks not tied to your local editor (the answer to Codex Ona and Devin Desktop).

If you have stability requirements, stay on 3.7 for a few more weeks. If you want the new SDK or Cloud Composer, opt into early-access from settings.

Windsurf SWE-1.5

SWE-1.5 is Windsurf’s current in-house coding model and the default for Cascade agents. The Windsurf pitch is hands-off long-running tasks: you describe a goal, Cascade plans, executes, and reports back, with optional human review at checkpoints.

Strengths:

Strongest hands-off autonomous workflow out of the four.
SWE-1.5 is well-tuned for refactoring and large structural changes.
Subscription includes generous Cascade quota.

Weaknesses:

Less editor polish than Cursor for in-line single-file work.
Less terminal-friendly than Claude Code for ad-hoc commands.
Smaller community/skill ecosystem than Claude Code’s Skills system.

Pricing: Pro tier $15/month is the cheapest of the four; higher tiers from $30 for power users.

Best for: Developers who want to delegate larger chunks of work and let the agent run.

Claude Code v2.1.169

The June 8, 2026 release. Notable additions:

—safe-mode flag (and CLAUDECODESAFEMODE env variable) starts Claude Code with all customizations off — no Skills, no MCP servers, no project CLAUDE.md, no agents directory.
/terminal-setup now disables GPU acceleration in VS Code, Cursor, and Windsurf integrated terminals to fix garbled-text rendering.
Feature-of-the-Week credit-claim status moved from a line above the prompt to a notification in the status area.
claude agents slash-command autocomplete now matches substrings.

The headline architectural features — background sub-agents, Skills, and Dynamic Workflows — are all production-ready in this build.

Pricing: Comes with any Claude Pro/Max plan ($20/$100/$200 respectively); Claude Fable 5 is on usage credits after June 22.

Best for: Developers who live in a terminal, want background autonomous sub-agents, or want the most mature Skills/MCP ecosystem.

Head-to-head decision table

You want	Pick
IDE-first workflow with strong tab completion	Cursor 3.7
Cloud Composer / hosted long-running tasks	Cursor 4 (early-access) or Codex Ona
Hands-off “describe goal, walk away” Cascade	Windsurf SWE-1.5
Terminal-first with autonomous sub-agents	Claude Code v2.1.169
Lowest price	Windsurf Pro at $15/mo
Best SWE-bench Pro score	Claude Code + Claude Fable 5
Best Terminal-Bench score	Codex CLI
Most mature MCP/Skills ecosystem	Claude Code

Benchmark snapshot (June 2026)

These numbers move weekly. As of mid-June:

SWE-bench Pro: Claude Fable 5 leading at ~69.2%, GPT-5.5 Pro close behind.
Terminal-Bench 2.1: Codex CLI on top, Claude Code second.
MCR (multi-file reasoning): Cursor 3.7 + Composer 2.5 strong on cached repos.

The model matters more than the agent for benchmark numbers. Claude Fable 5 powers strong results across Cursor, Windsurf, and Claude Code when you route to it. The differentiator between agents is workflow ergonomics, not raw model quality.

What about price?

Tool	Entry price	Power user tier
Cursor	$20/mo (Pro)	$40/mo (Business)
Windsurf	$15/mo (Pro)	$30+/mo
Claude Code	$20/mo (Claude Pro)	$200/mo (Max 20x)
Codex CLI	API metered	API metered

Cursor and Claude Code both cap at ~$200/mo for the power-user tier. Windsurf is the cheapest entry. Codex CLI is metered per-call against OpenAI’s API, so heavy users actually pay more than $200/mo despite no subscription floor.

What this means right now

If you started a project a month ago and want to upgrade: check whether you’re on Cursor 3.7 or earlier. Composer 2.5 is meaningfully faster than Composer 2.0 and worth the upgrade.

If you’re starting a new project this week: stay on Cursor 3.7 stable channel or pick Claude Code v2.1.169. Both are battle-tested with the current Claude Fable 5 / GPT-5.5 Pro generation. Skip Cursor 4 early-access unless you specifically want the new SDK.

If you want to delegate long-running work: Windsurf Cascade for editor-tied work, or Claude Code background sub-agents for terminal-tied work. Cursor 4’s Cloud Composer is the third option once it GAs.

If you live across multiple tools: Claude Code’s MCP servers can be invoked from inside Cursor and Windsurf via the MCP integration. So you don’t need to pick one — many developers use Cursor for the editor, Claude Code for autonomous terminal work, and Windsurf for long-running Cascade jobs in parallel.