If you opened a terminal to write code in 2026, you had to first choose which AI to put in it. Two years ago that meant Aider or a homegrown wrapper. Now there are half a dozen serious agentic CLIs, each with its own model lineup, pricing model, sandbox story, and quirks — and one of the big ones is being switched off in a few weeks. This is the roundup I wish someone had handed me: what each tool actually is, what it costs, whether you can read its source, and where it shines.
The contenders
The terminal AI coding space in 2026 splits into three camps. There are the vendor-locked frontier CLIs (Claude Code, Codex CLI, Antigravity CLI) tied to a lab's models. There's the open-source agentic tier (Qwen Code, Oh My Pi, OpenCode) you can read, fork, and point at any provider. And there's the old guard (Aider) that defined the category and is now coasting.
Claude Code (Anthropic)
The CLI software is free, but there's no standalone free tier — you bring a paid Claude subscription (Pro $20/mo, Max 5x $100/mo, Max 20x $200/mo) or pay-per-token API access. Anthropic's $0 plan covers the chat apps, not Claude Code.
As of June 2026 the default model is Claude Opus 4.8 (set as default with high effort, plus an /effort xhigh level, in Claude Code v2.1.154 on May 28, 2026). The lineup also includes Opus 4.7, Sonnet 4.6, Haiku 4.5, and the new Mythos-class Fable 5.
Usage is governed by a dual-layer system: a 5-hour rolling window plus a weekly cap on active compute hours. On May 6, 2026 Anthropic doubled the 5-hour limits for Pro/Max/Team and removed peak-hours throttling; weekly limits were raised 50% temporarily through July 13, 2026. It supports MCP, hooks, subagents, and skills, and Anthropic has added Managed Agents with self-hosted sandboxes (public beta), MCP tunnels (research preview), and a sandboxed Python code execution tool (public beta).
If your priority is raw code quality and multi-file reasoning, Claude Code is the one most 2026 comparisons rate best overall. The cost is being inside Anthropic's walled garden.
Codex CLI (OpenAI)
Codex CLI (@openai/codex) is the open-source counterweight — actively maintained, mostly Rust, with 88,600+ GitHub stars. Install with npm install -g @openai/codex or brew install --cask codex; it runs on macOS, Linux, and Windows via WSL. The CLI is free; you authenticate with a ChatGPT Plus/Pro/Business/Enterprise plan or bring your own OpenAI API key.
It's widely regarded as the strongest tool for sandboxing and token efficiency, scoring high on Terminal-Bench. It supports a configurable sandbox for running commands, AGENTS.md, skills in .codex/skills, MCP servers, and multi-step plans.
Pricing: the CLI is free software; usage is via a ChatGPT plan or BYOK. Plus is $20/mo; Pro starts at $100/mo where you pick a rate-limit multiplier at checkout (Pro 5x $100/mo, Pro 20x $200/mo). Starting April 2, 2026, Codex switched to token-based billing for Plus/Pro/Business, extended to Enterprise on April 23, 2026.
One point of confusion worth clearing up: the Codex CLI tool is not the same as the GPT-5 Codex models. The CLI runs on the GPT-5 family. Codex-specific models include GPT-5.1-Codex-Max (frontier long-running agentic coding), GPT-5.2-Codex, and GPT-5.3-Codex (Feb 5, 2026). Sources indicate the current CLI default is GPT-5.5 (April 23, 2026), while GPT-5.2-Codex remains the recommended model ID for API-key/BYOK workflows — the exact current default is genuinely ambiguous across sources, so check before you assume.
Antigravity CLI (Google) — and the Gemini CLI shutdown
This is the migration you can't ignore. Google's Gemini CLI is being deprecated. On June 18, 2026, Gemini CLI and the Gemini Code Assist IDE extensions stop serving requests for Google AI Pro/Ultra and free personal users. Existing installs show a deprecation warning until the deadline; afterward the auth endpoint returns 410 Gone and commands fail. Enterprise accounts can keep going via paid API keys.
The replacement is Antigravity CLI, built in Go as the successor to Gemini CLI. Antigravity launched May 19, 2026 as a four-surface platform: the Antigravity IDE, the Antigravity 2.0 desktop app, the Go-based CLI, and a Python SDK. It supports parallel execution and keeps Agent Skills, Hooks, Subagents, and Extensions (renamed plugins).
Pricing has a free tier with rate limits and no credit card, Pro at $20/mo, and Ultra at $249.99/mo. The free tier includes all models with rate limits. It supports multiple models including Gemini 3.1 Pro, Gemini 3 Flash, Claude Sonnet 4.6, Claude Opus 4.6, and GPT-OSS 120B. If you're on Gemini CLI today, treat June 18 as a hard deadline and migrate now.
Qwen Code (Alibaba)
Qwen Code is an open-source agent for the terminal, optimized for Qwen models but supporting OpenAI/Anthropic/Gemini-compatible APIs, Alibaba Cloud Coding Plan, OpenRouter, Fireworks AI, or BYOK. It ships Skills and SubAgents for a Claude Code-like workflow, and both the framework and the Qwen3-Coder model are open source.
Heads up on cost: the free OAuth tier was discontinued on April 15, 2026 — you now need API keys from a supported provider. Open-weight coding models include Qwen3-Coder-480B-A35B-Instruct (480B MoE, 35B active) and the newer Qwen3.6 series (Qwen3.6-27B scored ~71.78 Coding Avg on a May 12, 2026 LiveBench snapshot).
Oh My Pi (omp)
The most interesting newcomer. Oh My Pi is an MIT-licensed terminal coding agent by Can Bölük, forked from Mario Zechner's Pi. It's mostly TypeScript (Bun runtime, also Node) with some Rust, runs on macOS/Linux/Windows, and installs via curl -fsSL https://omp.sh/install | sh, brew install can1357/tap/omp, or bun install -g @oh-my-pi/pi-coding-agent. It's free; you bring your own model provider — 40+ of them (Anthropic, OpenAI, Gemini, xAI, Mistral, Groq, Ollama, etc.) plus MCP via an MCP Registry.
What sets it apart is engineering depth, not branding: hash-anchored ("Hashline") edits using two-character content-hash anchors plus line-numbered ops to slash output tokens (a reported 61% cut on Grok 4 Fast), full LSP integration so renames propagate, DAP-driven debugging against live binaries (lldb, dlv, debugpy), subagent orchestration across isolated worktrees, persistent Python/JS execution, and real browser automation. If Claude Code is the polished product, Oh My Pi is the tool that thinks like a senior dev about the editing loop itself.
The rest: Aider and OpenCode
Aider is the Apache-2.0, git-native pair-programming tool (~46k stars) that auto-commits per edit and uses a repomap, with local-model support via Ollama. As of mid-2026 its pace has slowed (last push ~May 22, 2026) and its model guidance hasn't been fully refreshed for 2026 frontier models — still useful, no longer cutting edge.
OpenCode is the MIT-licensed, provider-agnostic open-source agent that's become the de facto open-source Claude Code alternative. It's very actively developed (daily commits) and supports 75+ providers including OpenAI, Google, OpenRouter, and local models via Ollama or any OpenAI-compatible endpoint. Star counts vary by source (~150K–172K reported).
The comparison table
| CLI | Vendor | Source | Default / lead model | Pricing | Free / local option | MCP | Standout |
|---|---|---|---|---|---|---|---|
| Claude Code | Anthropic | Closed (free to use) | Opus 4.8 | Pro $20 / Max $100–$200; or API | No free tier | Yes | Best code quality, subagents, skills |
| Codex CLI | OpenAI | Open (Rust) | GPT-5.5 (ambiguous); GPT-5.2-Codex for BYOK | ChatGPT plan or BYOK; token-billed | BYOK only | Yes | Sandboxing, token efficiency |
| Antigravity CLI | Closed (Go) | Gemini 3.1 Pro (multi-model) | Free / Pro $20 / Ultra $249.99 | Free tier (rate-limited) | Yes | Multi-model, parallel exec | |
| Qwen Code | Alibaba | Open | Qwen3-Coder | BYOK (free OAuth ended Apr 15) | Local via providers | Yes | Open weights + framework |
| Oh My Pi | Can Bölük | Open (MIT) | BYO (40+ providers) | Free CLI, BYOK | Local via Ollama | Yes (Registry) | Hashline edits, LSP, DAP debug |
| Aider | Community | Open (Apache-2.0) | BYO | Free CLI, BYOK | Local via Ollama | Limited | Git-native auto-commit |
| OpenCode | Community | Open (MIT) | BYO (75+ providers) | Free CLI, BYOK | Local via Ollama | Yes | Best open-source all-rounder |
(Gemini CLI omitted — deprecated, stops serving June 18, 2026.)
How to choose
A few decision shortcuts that hold up in practice:
- Best output, money is no object: Claude Code with Opus 4.8.
- OpenAI shop or CI pipelines that run agent commands: Codex CLI — the sandbox and token efficiency pay off at scale.
- On Gemini CLI today: migrate to Antigravity CLI before June 18, 2026. Don't wait for the 410.
- Want to read and fork the source / avoid lock-in: OpenCode or Oh My Pi.
- Offline or cost-capped: any open CLI pointed at a local model via Ollama (GLM-5.1, Kimi K2.6, DeepSeek V4, Qwen3-Coder). Because every CLI here speaks the OpenAI-compatible API, a provider-agnostic, local-first routing layer like Wide Area AI can sit underneath them all — serving requests from your own hardware at zero per-token cost and failing over to cloud providers only when your local nodes are unavailable.
And the unglamorous truth from most 2026 teams: people don't pick one. They run Claude Code for hard refactors, Codex in CI, and an open CLI for local/offline work.
Bottom line
The terminal is where serious agentic coding lives in 2026, and the field has matured fast. Claude Code leads on quality but locks you in and has no free tier. Codex CLI is the open, sandbox-strong, token-efficient pick for OpenAI users. Antigravity is Google's mandatory upgrade path now that Gemini CLI is on a June 18 death clock. Qwen Code, Oh My Pi, and OpenCode give you open source, BYOK flexibility, and local-model freedom — with Oh My Pi pushing the most genuinely novel ideas about the edit loop. Match the tool to the job, keep an eye on the fast-moving model defaults and pricing, and migrate off Gemini CLI before it stops answering.