2026: бесплатные AI coding tools — полный token guide для Gemini, Claude, Copilot

Срез на 2026-06-09: free tier покрывает terminal agents, IDE plugins и open-source CLI aggregators. Таблица: tool → quota → paid required → EU access → strongest free lever.

Critical path: с 18 июня 2026 Google отключает Gemini CLI для free OAuth / Pro / Ultra users — migration на Antigravity CLI. Enterprise Gemini Code Assist и API key paths не затронуты. Разбор policy: Gemini CLI policy breakdown.

Bottom line: $0 coding stack 2026 = OAuth tools + Groq/OpenRouter free + open-source CLI aggregator. Три слоя закрывают большинство solo/small-team сценариев.

• Боль 1: subscription stack. Claude Pro $20, Cursor Pro $20, ChatGPT Plus $20, Copilot Pro $10 — full stack = $70+/month. Free tiers + student licenses закрывают ~60% daily tasks; подписываться на всё — hidden waste.
• Боль 2: OAuth vs BYOK. Gemini CLI и Codex CLI bind personal OAuth token к terminal session. Google explicitly запрещает routing через third-party proxy — ban risk. Compliance teams должны планировать API key / OpenRouter path.
• Боль 3: incomparable quotas. «1000/day», «2000 Tab/month», «14400 req/day» — разные unit of measure. Matrix в sec-3 мапит task intensity, не raw numbers.
• Боль 4: policy churn. Июнь 2026: Gemini CLI OAuth cutoff, Copilot usage-based с 1 июня. Single-vendor free tier = single point of failure — multi-tool backup + Groq/OpenRouter fallback.

Таблицы ниже и six-step checklist конвертируют боли в executable $0 path.

Quick routing: daily completions → Copilot Free or Cursor Hobby; terminal automation → Gemini/Antigravity or OpenCode + Groq; max reasoning → Claude Code with BYOK; multi-model gateway → OpenClaw. CLI rankings и Mac workflow mapping: OpenRouter CLI rankings.

1. Primary + backup tool: EU default stack: «OpenCode + Groq» as terminal agent, «Cursor Hobby» as IDE backup. With Google account: «Antigravity CLI» post-18.06. Students: activate GitHub Education for Copilot Student.
2. Grab free API keys: Groq (console.groq.com) — 14400 req/day; OpenRouter (openrouter.ai) — :free models; Google AI Studio — Gemini key. Set 80% quota alerts in dashboards.
3. Install OpenCode or Codex CLI: OpenCode: curl -fsSL https://opencode.ai/install | bash, then /connect. Codex: point ~/.codex/config.toml at OpenRouter or Groq (terminal block below).
4. Gemini CLI / Antigravity (optional): npm install -g @google/gemini-cli, Sign in with Google; migration plan to Antigravity before 18.06. Do not route OAuth token through third-party proxy.
5. IDE free tools: Cursor Hobby from cursor.com; Copilot in GitHub Settings → Copilot. Usage dashboard: app.cursor.sh/account/usage.
6. Validate first task + token habits: Flash models for trivial tasks; no /init full-repo scan; single-file prompts; OpenCode /models for switching. Log remaining quotas per platform.

openai_base_url = "https://openrouter.ai/api/v1"
model = "meta-llama/llama-3.3-70b-instruct:free"
sandbox_mode = "workspace-write"
web_search = "disabled"
approval_policy = "on-request"

export OPENROUTER_API_KEY="sk-or-v1-xxxxxxxx"
export OPENAI_API_BASE="https://openrouter.ai/api/v1"
export GROQ_API_KEY="gsk_xxxxxxxx"

OpenAI-compatible endpoints, plug into OpenCode, Codex CLI, OpenClaw:

• Groq: https://api.groq.com/openai/v1; 14400 req/day free; Llama 3.3, Mixtral, minimal latency.
• OpenRouter: https://openrouter.ai/api/v1; :free model suffix; 500+ models, zero markup BYOK on many tiers.
• Google AI Studio: Gemini REST API; free tier with rate limits; post-Antigravity key path.
• Ollama local: ollama pull qwen3:8b → OpenCode/OpenClaw; zero API burn, needs RAM/GPU.

Citable data (vendor docs, Google Developers Blog, 2026-06-09):

• Gemini CLI OAuth: 60 RPM, 1000 RPD; Gemini 2.5 Pro OAuth path ~100 RPD, 5 RPM.
• Antigravity CLI: personal preview free from June 2026; default Gemini 3.5 Flash; absorbs free/Pro users after 18.06.
• Cursor Hobby: 2000 Tab/month + 50 slow premium agent requests; student Pro 1 year via GitHub Education.
• GitHub Copilot Free: 2000 completions + 50 premium/month; Student ≈ Pro, 300 premium/month, value $10/month.
• Groq Free: 14400 requests/day; Llama 3.3 70B and Mixtral available.
• OpenCode: 146K+ GitHub stars; 75+ native providers.

Token-saving hacks (extract):

1. No /init full-repo scan — Claude Code and Codex burn context aggressively.
2. Single-file scope only, never «optimize entire project».
3. Trivial tasks: Gemini 2.5 Flash-Lite or Groq Llama; complex: Pro tier.
4. 80% quota alert in Groq/OpenRouter/Google Console.
5. Ollama + OpenCode for repetitive jobs — zero API cost.
6. Short system prompts — less per-request overhead.
7. OpenClaw models scan for OpenRouter :free models.
8. Multi-key rotation against 429 rate limits.

Free tokens решают inference cost, не machine uptime. OpenClaw Gateway, Codex sandbox, Claude Code headless CI = 24/7 requirement. На laptop или oversubscribed VPS типичные failure modes:

• Network jitter: SSH drop, OAuth refresh timeout, agent task dies mid-flight.
• VPS oversubscription: Docker sandbox + sub-agents fight for CPU — OOM killer.
• Laptop sleep: launchd gateway stops after lid close — no 24/7.

Teams running multi-CLI + Ollama + iOS builds in parallel: free API не заменяет production host. JEXCLOUD multi-region bare-metal Mac — dedicated Apple Silicon, real macOS, no overselling, ~120s provisioning, monthly elastic rental; tokens stay Groq/OpenRouter BYOK, machine and inference decoupled. Claude Code Seatbelt sandbox, Goose Docker, Xcode build chain — most stable on bare metal.

Sizing cheat sheet: light OpenCode + Groq → MacBook Air M3 16GB; heavy OpenClaw + Ollama → Mac mini M4 Pro 32GB+; local 14B inference → Mac Studio 64GB+ unified memory. Hackathon/MVP: rent beats buy. Specs: JEXCLOUD pricing; remote access: help center.