A sequel to my most-read post. Six months later, the landscape exploded. Here's every tool that matters, what it costs, and what it actually does.
The AI coding tool market went from "a few options" to "overwhelming" in about six months. New CLIs weekly. Pricing wars. Open-source alternatives rivaling the paid ones. Chinese models hitting 77%+ on SWE-bench. Free tiers that would've been unthinkable a year ago.
I've tested, researched, or tracked 30+ tools. Here's the complete map.
Tier 1: Cloud Subscriptions (Pay Monthly, They Host Everything)
These are the "just works" options. You pay, they handle models, infrastructure, and updates.
| Tool | Monthly Cost | Model(s) | Type | SWE-bench | Notable |
|---|---|---|---|---|---|
| Claude Code | $17-20 (Pro), $100-200 (Max) | Claude 4.6 Opus/Sonnet | Terminal agent | 80.9% | 1M context. Uses 5.5x fewer tokens than Cursor. Full hook/plugin system. |
| Cursor | $16/mo | Multi-model | VS Code fork | Varies | Largest community. Best tab completions. Most polished UX. |
| Windsurf | $20/mo | Multi-model | IDE | Varies | "Flows" persistent context. Raised from $15 in March 2026. |
| Codex CLI | With ChatGPT Plus ($20/mo) | GPT-5 series | CLI + Desktop | — | Cloud sandbox execution. Autonomous agent. |
| Antigravity | $20 (Pro), $250 (Ultra) | Gemini | Agent IDE | — | Google's entry. Parallel agents. Built-in Chrome for testing. |
| Mistral Vibe | $15/mo (Le Chat Pro) | Devstral 2 | CLI | — | Apache 2.0 source code. Paid models. |
| Amp (Sourcegraph) | Free tier ($10/day cap) | Multi-model | CLI + IDE | — | "Deep mode" autonomous research. No markup on API costs. |
The verdict: Claude Code wins on capability (1M context, best SWE-bench, hook system). Cursor wins on UX. Windsurf and Antigravity bet on parallel agents. Codex bets on cloud sandboxing.
Token efficiency matters more than subscription price. Claude Code using 5.5x fewer tokens than Cursor means the real cost difference is bigger than the $1-4/mo subscription gap suggests.
Tier 2: Genuinely Free (Real Usage, No Tricks)
These tools offer meaningful free access — not "free trial" but actually usable for daily work:
| Tool | Free Tier | What You Get | Upgrade Path |
|---|---|---|---|
| Gemini CLI | 1,000 requests/day | Gemini 2.5 Pro/Flash routing. Just login with Google. | Pay-as-you-go |
| GitHub Copilot CLI | 50 premium requests/mo | Deep GitHub integration. Natural for existing users. | $10/mo |
| Amazon Q Developer | Free tier | Best for AWS-heavy workflows. | AWS pricing |
| Kiro (Amazon) | Free tier | Spec-driven: generates requirements before code. Auditable trail. | TBD |
| Qwen Code | Free API (!) | Alibaba's CLI agent. Apache 2.0. Completely free API access. | — |
Gemini CLI at 1,000 free requests/day is the story here. For many developers, this is effectively unlimited. If you're budget-constrained or evaluating, start here.
Qwen Code's free API is underappreciated. Alibaba is subsidizing it for market share — take advantage while it lasts.
Tier 3: Open Source BYOK (Free Tool, Bring Your API Key)
The largest category. Zero subscription — you pay only for model inference via API keys:
| Tool | GitHub Stars | Type | Model Support | What Makes It Different |
|---|---|---|---|---|
| OpenCode | 140K+ | CLI | 75+ providers | Universal adapter. If a model exists, OpenCode supports it. |
| Aider | 39K+ | CLI | Any (inc. local) | Git-native. Auto-commits. Most mature. 4.1M installs, 15B tokens/week. |
| Cline | — (5M installs) | VS Code ext | Any | Most adopted open-source coding extension. |
| Continue.dev | 26K | IDE ext | Any | Only tool with full VS Code + JetBrains support. |
| Goose | — | CLI + Desktop | Any + MCP | Block/Square's agent. Apache 2.0. Native MCP integration. |
| Roo Code | — | VS Code ext | Any | "When other agents break down" — reputation for reliability on large multi-file changes. |
| OpenClaw | — | CLI | GLM, MiniMax, Qwen, etc | Gateway to Chinese model ecosystem. |
| Zed | — | Editor | BYOK | Rust-native. Fastest editor in the category. |
| iFlow | — | CLI | Any OpenAI-compatible | SubAgents. Controlled file permissions. |
| Kimi Code CLI | — | CLI | Kimi K2.5 | Moonshot's agent. 100-agent swarm capability. |
| BLACKBOX | — | Multi | Proprietary + BYOK | Completions + chat + search. |
The real cost of BYOK: With Claude Sonnet at $3/$15 per million tokens, moderate daily use runs $10-15/month. With OpenRouter, you can compare prices across 100+ models. With local models, the cost is $0.
Aider remains the gold standard for terminal pair-programming. Git-native workflows, clean commit history, works with everything from GPT to local Ollama models.
Tier 4: Truly Local (Offline, Self-Hosted, Zero Cloud)
For the privacy-conscious, air-gapped environments, or anyone who wants zero recurring costs:
Inference Runtimes
| Runtime | Best For | Effort | Speed |
|---|---|---|---|
| Ollama | Easiest start. One command: ollama pull qwen2.5-coder
|
Minimal | Good |
| llama.cpp | Maximum control. Custom compilation for your exact hardware. | High | Best (tuned) |
| LM Studio | Visual model management. Side-by-side comparison. GUI sliders. | Minimal | Good |
| vLLM | Production serving. PagedAttention cuts memory 50%+. 2-4x throughput. | Medium | Production-grade |
| Tabby | Self-hosted copilot. Full IDE integration on your own infra. | Medium | Good |
Best Local Coding Models (April 2026)
| Model | Params | SWE-bench | License | Runs On |
|---|---|---|---|---|
| GLM-5 (Zhipu) | 744B MoE (40B active) | 77.8% | MIT | vLLM / llama.cpp (needs 80GB+ VRAM for full) |
| Kimi K2.5 (Moonshot) | 1T MoE | 76.8% | Open | Similar — enterprise hardware |
| Devstral 2 (Mistral) | — | — | Apache 2.0 | Ollama, llama.cpp |
| Qwen 2.5 Coder (Alibaba) | 7B-72B | — | Apache 2.0 | Ollama (7B on laptop, 32B on desktop) |
| MiniMax M2 | 230B MoE (10B active) | — | Open | 8% of Claude's price, 2x speed |
| DeepSeek Coder V2 | Various | — | MIT | Ollama, llama.cpp |
For a laptop: Qwen 2.5 Coder 7B or DeepSeek Coder V2 7B via Ollama. Runs fine on 16GB RAM.
For a desktop with GPU: Qwen 2.5 Coder 32B via Ollama. Excellent quality, runs on RTX 3060 12GB.
For a server: GLM-5 or Kimi K2.5 via vLLM. These compete with Claude on coding benchmarks.
Tier 5: Model Routers (Connect Anything to Anything)
| Router | What It Does |
|---|---|
| 9router | Connects 40+ providers to Claude Code, Cursor, Copilot, Antigravity, etc. |
| CLIProxyAPI | Wraps Gemini CLI, Codex, Claude Code as OpenAI-compatible API. Use free Gemini models through any tool. |
| OpenRouter | Universal API gateway. Compare prices across 100+ models. Pay-per-token. |
CLIProxyAPI is wild: it wraps Gemini CLI's free tier as an OpenAI-compatible API, which means you can use Gemini 2.5 Pro through Aider, Cline, or any OpenAI-compatible tool — for free.
Quick Decision Matrix
| If you want... | Use this |
|---|---|
| Best capability, cost be damned | Claude Code (Max) |
| Best free experience | Gemini CLI |
| Best open-source CLI | Aider |
| Best IDE experience | Cursor |
| Best for teams | Continue.dev (VS Code + JetBrains) |
| Zero cloud dependency | Ollama + Qwen 2.5 Coder |
| Best Chinese model access | OpenClaw |
| Planning before coding | Kiro |
| Git-native workflows | Aider |
| Parallel agents | Antigravity or Windsurf |
Next in this series: *Part 2 — Running AI Coding Agents for Free: The Open Source & Local Guide** — deep dive into BYOK setups, local model configuration, and getting Claude-level performance without a subscription.*
Also: *Part 3 — What Every AI Coding Tool Gets Wrong** — the measurement gap that none of these tools address.*
This is a sequel to The best (free - cheap) AI friendly Cli and Coding environments.
Top comments (0)