ИЮНЬ 2026 MAC
БЕСПЛАТНЫЕ_
AI_ТОКЕНЫ_
КОДИНГ.
В июне 2026 плотность бесплатных AI-токенов для кодинга — максимум, но Gemini CLI cutoff 18.06 уже на горизонте. Боль: Mac-разработчики крутят Cursor, Copilot, Gemini CLI и OpenCode, не зная, что ещё $0, где нужен proxy и чем заменить Gemini после cutoff. Вывод: одна quota-матрица, пять шагов деплоя и трёхуровневая routing-таблица Mac дают daily agent workflow без кредитки; CN-dev могут подключить SiliconFlow/Bailian/Zhipu без proxy в Codex CLI и OpenCode. Структура: quota table → pain points → Gemini migration → domestic APIs → IDE free tiers → terminal agents → five steps → case study → Mac rental decision.
1. Pain points: free tier много, правда фрагментирована
1) Policy window узкий: Gemini CLI OAuth ещё даёт 1 000 req/day @ 60 RPM, но Free/Pro/Ultra personal теряют доступ 18.06 — official path: closed-source Antigravity CLI (см. наш разбор политики Gemini CLI). 2) Network split: Google OAuth из mainland CN требует proxy; SiliconFlow — 20M permanent tokens, Alibaba Bailian — 70M tokens, Zhipu — 20M tokens, всё direct. 3) IDE vs CLI quotas не стакаются: Cursor Hobby — 2 000 Tab completions + 50 slow premium req/month; Copilot Free зеркалит — single-tool budget fail; побеждает multi-tool rotation. 4) Token sinkholes: Claude Code и Codex /init сканируют весь repo — сотни тысяч tokens за один shot. Free tier: ban pattern. 5) Metal throughput: на M2 Air 16 GB параллельный Ollama Qwen3:8b + Gemini stream + Docker sandbox часто упирается в unified memory до того, как cloud quota кончится — decode tok/s падает с ~45 до <15 при UM >85 %.
2. Одна таблица: snapshot free quota — июнь 2026
| Tool / Platform | Free Quota | Paid Account? | CN Direct | Best Leverage |
|---|---|---|---|---|
| Gemini CLI (OAuth) | 1 000 req/day, 60 RPM | No — Google account | Proxy needed | Last window before 6/18, no card |
| Gemini API (AI Studio) | Flash 1 500/day | No | Proxy needed | 2.5 Pro 100/day |
| Codex CLI | ChatGPT free (limited) | No | No | Point to domestic API |
| OpenCode | Tool free + BYOK | No | Yes with domestic API | 75+ providers, 146K stars |
| OpenClaw | Tool free + multi-provider | No | Yes | Reuse Gemini OAuth / domestic keys |
| GitHub Copilot Free | 2 000 completions + 50 premium/mo | No | Yes | Enable from GitHub settings |
| Cursor Hobby | 2 000 Tab + 50 slow premium/mo | No | Yes | Full VS Code fork |
| SiliconFlow | 20M tokens (permanent) | No — CN phone | Yes | DeepSeek / Qwen / GLM-5 |
| Alibaba Bailian | 70M tokens | No — Alipay KYC | Yes | 70+ models |
| Zhipu AI | 20M tokens (permanent) | No | Yes | GLM-5, Claude relay |
Back-of-envelope: Gemini CLI (pre-cutoff) + SiliconFlow + Cursor Hobby + Copilot Free экономит ~$30–$80/month vs paid stack — хватит на несколько дней rented remote Mac node. Логируйте baseline: tokens/run, p95 TTFT, decode tok/s, UM peak GB.
3. Gemini CLI: последние дни до 18.06
Post-cutoff paths: (1) Antigravity CLI (agy, Google account, tighter limits); (2) Gemini API Key via AI Studio free tier; (3) SiliconFlow + OpenCode/Codex без Google OAuth. Routing Gemini CLI OAuth через third-party proxy — ToS violation, account suspension risk.
4. No-VPN CN route: SiliconFlow / Bailian / Zhipu
Stack recommendation: daily coding — DeepSeek-V3 (SiliconFlow); long context — Qwen3.5-Max (Bailian); Claude Code relay — Zhipu GLM-5. 80% quota alerts в каждой console. Для local draft loop: Ollama Qwen3:8b на Metal — ~38–52 decode tok/s на M3 Pro при 8K context, UM ~6 GB.
5. IDE free tiers: Cursor Hobby и GitHub Copilot
| Feature | Cursor Hobby | Copilot Free | Copilot Student |
|---|---|---|---|
| Completions | 2 000/month | 2 000/month | Unlimited |
| Premium agent | 50 slow/month | 50/month | 300/month (~Pro) |
| Student path | cursor.com/students — 1yr Pro | GitHub Education | .edu email |
Rotation: Tab completions на Copilot (preserve Cursor slow quota), refactors на Cursor Agent (50 wisely), batch terminal на OpenCode + domestic API. Students: GitHub Education + Cursor Student Pro — ещё $240/year saved.
6. Terminal agent stack: OpenCode и OpenClaw
OpenCode — multi-model experiments; OpenClaw — 24/7 IM automation. Оба tools free — token cost = attached API. Domestic free tiers = $0 run rate. Gateway на remote Mac через launchd держит stable Metal throughput без laptop thermal throttle.
7. Пять шагов: zero → running free agent workflow
Step 1 — Register domestic API, set quota alerts
SiliconFlow, Bailian или Zhipu; два provider для failover. Keys в ~/.zshrc, never commit to Git.
Step 2 — Install OpenCode или Codex CLI, verify endpoint
codex doctor или OpenCode /connect; single-file prompt; latency <3s; log p95 TTFT.
Step 3 — Gemini CLI OAuth до 18.06 (optional)
С proxy — burn 1 000/day window; export AI Studio API key backup.
Step 4 — Configure Cursor + Copilot dual-IDE rotation
Cursor custom model → SiliconFlow OpenAI endpoint; Copilot owns Tab; Cursor slow для hard refactors.
Step 5 — Model switch cheat sheet и token discipline
Flash/Lite easy tasks, Pro hard; never /init full repo; single-file ops; local Ollama Qwen3:8b drafts @ Metal sidecar.
8. Десять token-saving rules (free tier mandatory)
Ban /init repo scans. Single-file scope only. Gemini Flash-Lite (1 500/day) drafts. 80% quota alerts. Separate Google accounts for independent Gemini API pools (within ToS). Ollama + OpenCode local drafts — measure decode tok/s before cloud call. Short system prompts. OpenRouter free models via OpenClaw scan. Multi-key load balance vs 429s. Watch vendor promo credits. Log UM peak GB alongside token bills — thermal throttling = silent quota on Apple Silicon.
9. Mac three-tier routing matrix
| Tier | User | Stack | Hardware / Throughput |
|---|---|---|---|
| Local light | Student, side projects | Cursor Hobby + SiliconFlow + Gemini CLI (until 6/18) | MacBook Air 16GB — Ollama draft ~35 tok/s |
| Multi-tool rotation | Indie dev | OpenCode + Copilot Free + dual API failover | MacBook Pro 24–32GB — decode stable >40 tok/s |
| Remote always-on | 24/7 agent / team | OpenClaw Gateway + launchd + API pool | MACGPU Mac mini M4 Pro — 7×24 Metal queue |
10. Case study: indie dev saves $67/month
«Solo dev в Hangzhou платил Cursor Pro ($20/mo) + Claude API (~$47/mo), 16GB MacBook Air swap'ил под OpenClaw — decode tok/s падал с 48 до 12 при UM 92 %. После restructure: Copilot Free Tab, Cursor 50 slow + SiliconFlow DeepSeek-V3 refactors, Gemini CLI scripts до cutoff, OpenClaw на MACGPU M4 Pro 32GB remote. Month-end: $0 subscriptions, 18M SiliconFlow tokens left — $67 equivalent saved. Local Mac — Cursor review + Final Cut; agents cron на remote Mac overnight @ stable Metal throughput.»
Industry read: mid-2026 free tiers — не demos, а production validation. Но policy windows сужаются (Gemini CLI cutoff, partial Copilot Pro signup pauses с April 2026). Smart play: prove workflows на free tiers, offload 24/7 Gateway на remote Mac вместо pin laptop UM и thermals к Docker sandboxes. Apple Silicon unified memory — best price/performance base для hybrid local-Ollama-draft + cloud-API-primary; offload batches на 128 GB remote когда M2/M3 peaks >80 % UM during agent+ComfyUI overlap.
11. Citable numbers и checklist
Gemini CLI OAuth: 1 000 req/day, June 18 cutoff. SiliconFlow: 20M tokens permanent. Bailian: 70M tokens. Cursor Hobby: 2 000 Tab + 50 slow/mo. Copilot Student: ~$10/mo value. OpenCode: 146K+ GitHub stars. Local Ollama Qwen3:8b @ M3 Pro: ~45 decode tok/s p50.
Checklist: Domestic API configured □ | Codex/OpenCode single-file test pass □ | Gemini OAuth or migration plan □ | Cursor/Copilot rotation documented □ | /init ban in team policy □ | 80% quota alert set □ | Mac three-tier routing labeled □ | Metal tok/s baseline logged □
Windows и Linux тоже run SiliconFlow и OpenCode, но macOS wins для parallel Xcode/Final Cut/ComfyUI workflows, Claude Code Seatbelt sandboxing, launchd 24/7 OpenClaw, Metal-sidecar Ollama validation. Если free-agent stack на 16GB Mac trigger thermal throttling или Docker memory pressure — MACGPU remote Mac nodes (M3 Pro 32GB / Mac mini M4 Pro) host OpenClaw/Hermes и headless CI; laptop keeps Cursor review и light API calls. Predictable monthly rent vs stable throughput — log fan RPM и decode tok/s weekly; rebalance когда cloud p95 TTFT regresses >15 % при steady Metal remote.