Июнь 2026: сравнение AI-кодинг-помощников для Mac — dual-stack

Июнь 2026: AI-кодинг-инструменты — это не LSP-плагины, а автономные coding agents с multi-file edit, shell execution и PR pipeline. Технический тезис: SWE-bench Verified у Claude Opus 4.7 — 87.6%; Copilot Agent Mode — ~56%. С 1 июня 2026 Copilot биллит в AI credits (1 credit = $0.01). Оптимальный Mac-стек: Cursor (IDE layer) + Claude Code (agent runtime). Copilot — для GitHub-centric enterprise; Gemini/Antigravity — в миграционном limbo. Ниже: runtime-архитектуры, token economics, SWE-bench breakdown, пятиступенчатый deployment runbook.

1. Root cause analysis: три слоя неопределённости

Layer 1 — Benchmark ≠ production throughput. SWE-bench Verified измеряет fix-rate на изолированных GitHub issues; ваш monorepo с legacy DI, custom linters и 40K LOC test suite даёт иной профиль latency/accuracy. Spread 87.6% → 56% — не повод слепо брать #1.
Layer 2 — Billing model shift. Copilot с 01.06.2026: agent invocations + large-context requests жрут credits; inline completion — cost = 0. Cursor: dual-pool — Auto/Composer (включён в Pro $20) vs third-party API routing (отдельный burn rate).
Layer 3 — Runtime deprecation. Gemini CLI EOL 18.06.2026 для Free/Pro/Ultra → миграция на Antigravity CLI (agy). Детали: политика Gemini CLI.
Layer 4 — Unified Memory pressure. Cursor + Docker sandbox + Ollama на 16GB M-series → swap thrashing. Agent workloads offload на remote node (Cursor Agent Skills).

2. Runtime taxonomy: IDE fork vs headless agent

Runtime	Vendor	Execution model	Primary interface
Cursor	Cursor Inc.	Electron IDE, VS Code fork	GUI + inline Composer agent
Claude Code	Anthropic	Headless CLI agent	`claude` REPL / `claude -p`
GitHub Copilot	Microsoft	LSP extension + agent service	VS Code, JetBrains, Neovim, …
Antigravity	Google	Go CLI, async job queue	`agy` terminal / desktop

3. Cursor 3.5: Composer 2.5 internals

Composer 2.5 (May 2026, fine-tune на Kimi K2.5): multi-file diff engine, context window до ~256K tokens в production routing. Cloud Agents: isolated VM (не ваш local Metal), async multi-repo, push PR через GitHub App. BugBot: static + LLM review на PR diff.
Pricing: Pro $20/mo → $20 credit pool; Auto mode = unlimited в рамках fair-use. Team Standard $40/seat/mo с July 2026.
SWE-bench Multilingual: 73.7% (Composer 2.5 checkpoint).

# Cursor: типичный Composer session
# Context: @codebase + @docs + 12 files pinned
# Agent loop: plan → edit → terminal → verify

cursor --composer "refactor auth middleware to JWT RS256"
# Cloud Agent (async):
# Settings → Cloud Agents → repo: org/backend → branch: feat/jwt
                

Pros: lowest migration cost с VS Code; Tab completion latency <100ms на M-series; visual diff UX. Cons: Composer SWE-bench ниже Claude Code; Cloud Agent extra billing; vendor lock-in на fork.

4. Claude Code: agent loop & 1M context window

Headless agent runtime. Claude Opus 4.7: 1,000,000 token context (effective для whole-repo indexing через chunked retrieval), SWE-bench Verified 87.6% — industry peak, April 2026. GitHub: 110K+ stars.
Core loop: Explore → Plan → Implement → Commit. Plan Mode: read-only, no file writes — safe для architecture review. Agent Teams: spawn sub-agents с isolated context. CLAUDE.md: persistent system prompt injection per project root. MCP: tool protocol для DB, browser, custom APIs.

# Install & init
npm install -g @anthropic-ai/claude-code
cd ~/projects/backend && claude

# Plan Mode (read-only architecture pass)
claude --plan "map dependency graph of /src/auth"

# Programmatic invocation (CI/CD)
claude -p "fix failing test in auth_test.go" --output-format json

# CLAUDE.md snippet (project root)
## Build: make test && make lint
## Never modify: /migrations/*
## Commit format: conventional commits
                

Pricing: Pro $20/mo; Max 5x $100/mo (recommended для daily agent workloads); Max 20x $200/mo. API calls (claude -p, GitHub Actions) — per-token billing отдельно.
Pros: highest SWE-bench; 1M context; IDE-agnostic; Seatbelt sandbox на macOS. Cons: no Tab completion; Claude-only models; steep CLI learning curve; terminal-only UX.

5. GitHub Copilot: credit metering с 1 июня

Credit model (effective 2026-06-01): 1 credit = $0.01 USD. Pro $10/mo → 1500 credits ($15 value); Business $19/seat/mo; Enterprise $39/seat/mo.
Critical: inline code completion — credits_consumed = 0. Agent Mode, chat with large context, Workspace automation — metered.
Model routing: OpenAI, Anthropic, Google, xAI — switchable per session. Copilot Workspace: Issue → branch → PR pipeline.

# Copilot credit burn profile (approximate)
inline_completion     → 0 credits/request
agent_mode_multi_file   → 5-50 credits/task (context-dependent)
workspace_issue_to_pr   → 20-100 credits/issue

# SWE-bench Agent score: ~56%
# Enterprise: SSO, audit logs, policy enforcement — maturest stack
                

4.7M+ subscribers; 90% Fortune 100. Optimal: GitHub-native teams, budget entry $10, multi-IDE fleets.

6. Gemini → Antigravity: migration path

Gemini CLI (gemini) → Antigravity CLI (agy, Go rewrite, async background workers). 2026-06-18: shutdown для Free/Pro/Ultra personal tiers. Enterprise Code Assist — unchanged.
Gemini 3.1 Pro SWE-bench Verified: 80.6%. Multimodal input (code + image + PDF) — unique capability. Open-source Gemini CLI (Apache 2.0): security patches only, no feature commits.

# Migration checklist (personal tier)
gemini --version          # check current
agy auth login            # Antigravity OAuth
agy config set model gemini-3.1-pro

# Risk: feature parity not 1:1 as of June 2026
# Enterprise Code Assist: no action required
                

7. Full comparison matrix

Parameter	Cursor	Claude Code	Copilot	Gemini/Antigravity
Min paid tier	Pro $20/mo	Pro $20/mo	Pro $10/mo	transition
Recommended solo	Pro $20/mo	Max 5x $100/mo	Pro $10/mo	TBD
Max context	~256K	1M tokens	up to 1M (credit-heavy)	model-dependent
Tab completion	excellent	N/A	excellent (unlimited, 0 credits)	available
Multi-file agent	excellent	strongest	good	good
Model providers	multi	Claude only	4 vendors	Gemini only
IDE support	own IDE	any (CLI)	7+ editors	VS Code/JetBrains/CLI
SWE-bench Verified	73.7% (Composer)	87.6%	~56%	80.6% (Gemini 3.1 Pro)
macOS sandbox	Electron sandbox	Seatbelt	extension sandbox	CLI process isolation

SWE-bench Verified ranking (April 2026)

Claude Opus 4.7 (Claude Code)    ████████████████████████ 87.6%  ← #1 industry
GPT-5.3-Codex                    ███████████████████████  85.0%
Claude Opus 4.6                  █████████████████████    80.8%
Gemini 3.1 Pro                   █████████████████████    80.6%
Cursor Composer 2.5              ███████████████████      73.7%
GitHub Copilot Agent             ██████████████           56.0%
                

8. Пятиступенчатый dual-stack deployment runbook

Step 1 — Task routing matrix. Tab + micro-edits → Cursor/Copilot (0 marginal cost на completion). Cross-repo refactor 10+ files → claude --plan then implement. Issue→PR → Copilot Workspace или Cursor Cloud Agent.
Step 2 — Budget tiering. Entry: Copilot Pro $10. Standard dual-stack: Cursor Pro + Claude Pro = $40/mo. Heavy: Claude Max 5x + Cursor Pro = $120/mo.
Step 3 — Config sync. Maintain CLAUDE.md + .cursor/rules с identical style constraints. Drift = inconsistent codegen output.
Step 4 — Credit/token monitoring. Copilot: set monthly credit cap в org settings. Cursor: track Auto pool vs API pool separately. Alert at 80% threshold.
Step 5 — Mac three-tier compute routing. Tier A (local): Cursor editing only. Tier B (remote Mac): Claude Code long-running agents via SSH. Tier C (local MLX): draft validation на mlx_lm.server перед API burn.

# Step 5: remote Claude Code via SSH (MACGPU node)
ssh macgpu-node 'cd /workspace/backend && claude -p "migrate ORM layer"'

# Local: только review
git fetch origin && git diff origin/main...HEAD

# launchd 7x24 agent (remote node)
# ~/Library/LaunchAgents/com.macgpu.claude-nightly.plist
                

9. Scenario decision matrix

Workload	Tool	Technical rationale
Daily multi-file editing	Cursor Pro	Best diff UX, lowest context-switch cost
Architecture refactor	Claude Code Max	87.6% SWE-bench, 1M context, Plan Mode
Enterprise compliance	Copilot Business	SSO, audit logs, $19/seat
Budget solo dev	Copilot Pro	$10/mo, completion at 0 credits
Google Cloud native	Antigravity CLI	GCP/Workspace integration
Cross-repo automation	Cursor Cloud Agent	Isolated VM, parallel repos

10. Case study: 10-person Mac team — metrics before/after

«SaaS team, 10 Mac devs, baseline: all Cursor Pro ($200/mo) + Claude API overages (~$380/mo avg). Re-architected per runbook: ① Copilot Business for Tab completion ($190/mo, credits ≈ 0 on inline); ② 3 seniors on Claude Code Max 5x ($300/mo); ③ 2× MACGPU M4 Pro 32GB remote nodes for nightly claude -p migration batches. 3-month delta: SWE-bench-class task duration -42%, API overages $0, 16GB Air swap pressure eliminated. Total: $490/mo vs $580+ prior. p99 agent latency on remote node: 340ms vs 2.1s local under memory pressure.»

OpenRouter weekly token data Q2 2026: CLI agents (Claude Code, Hermes) — sustained growth. См. OpenRouter CLI ranking. Billing reforms делают «full Ultra everywhere» economically irrational. Split interactive editing (low burn) from autonomous agents (high burn); route long jobs to remote Mac 7×24.

11. Reference constants & acceptance checklist

① Claude Opus 4.7 SWE-bench Verified: 87.6%. ② Cursor Composer 2.5 Multilingual: 73.7%. ③ Copilot Pro entry: $10/mo. ④ Copilot credit rate: 1 = $0.01 (from 2026-06-01). ⑤ Claude Max 5x: $100/mo. ⑥ Gemini CLI personal EOL: 2026-06-18.

Windows/Linux run Copilot + Claude Code CLI, но macOS выигрывает на Xcode sidecar, Seatbelt sandbox, launchd persistent agents, Metal MLX local draft loop. Если 16GB unified memory saturated agent workloads — MACGPU remote Mac (M3 Pro 32GB / Mac mini M4 Pro) dedicated для Claude Code long-running tasks; local machine — только Cursor interactive editing + diff review.