The three agentic coding CLIs that matter in mid-2026 — Anthropic's Claude Code, OpenAI's Codex CLI, and GitHub Copilot CLI. All three sit in the terminal, read codebases, plan multi-file changes, run tests, and iterate on failures. They differ in model lock-in, sandbox model, pricing shape, and extensibility. This leaf maps where each wins.
All three operate at the project level rather than line-by-line autocomplete. Read codebases, plan changes, edit multiple files, run tests, iterate. The differences begin at the model layer.
The same engine the Claude Agent SDK runs on. Locked to Claude (Sonnet, Opus). Native Skills, hooks, MCP, sub-agents, plan-mode. Bundled with Claude Pro / Max / Team / Enterprise subscriptions.
Open-source CLI binary (Rust, MIT) backed by OpenAI's codex-1 (o3-based) and codex-mini (o4-mini-based) models. Cloud-sandboxed by default; parallel task execution. Bundled with ChatGPT Plus / Pro / Business / Enterprise.
Routes between Claude (Sonnet/Opus 4.6) and GPT-5.2–5.5 from one CLI; switch with /model. Worktree-isolation, self-iterating code review, Plan vs Autopilot modes. Bundled with Copilot Pro / Pro+ / Business / Enterprise; moves to usage-based AI Credits June 1 2026.
Where the three differ in shape. "Different" is more common than "better."
| Dimension | Claude Code | Codex CLI | Copilot CLI |
|---|---|---|---|
| Maker | Anthropic | OpenAI | Microsoft / GitHub |
| CLI license | Proprietary | Open source (Rust) | Proprietary |
| Model | Claude only (Sonnet, Opus) | OpenAI only (codex-1, codex-mini) | Multi-model — Claude 4.5/4.6 + GPT-5.2–5.5 |
| Switch models | n/a (single model) | n/a (OpenAI-only) | /model mid-task |
| Where it runs | Terminal · VS Code · JetBrains · claude.ai/code web | Terminal · web (codex.app) · IDE · Chrome extension | Terminal · VS Code · JetBrains delegation · Copilot Workspace |
| Sandbox model | Local execution; approval-gated | Cloud sandboxes by default; parallel tasks | Worktree or workspace isolation; local |
| Approval model | Approval-by-default; Plan + Autopilot | Approval-by-default | Plan (approval) or Autopilot (autonomous) |
| Sub-agents | Yes — context-isolated with restricted tools | Code-review agent built in | Yes — Explore, Task + custom in ~/.copilot/agents/ |
| MCP | Native; first-class | Native | Native |
| Skills / customisation | Skills, hooks, sub-agents, slash commands | Modal Vim editing, custom workflows, MCP | Custom agents, /delegate, image analysis |
| IDE bridge | VS Code + JetBrains extensions | Various IDE integrations | JetBrains delegation (May 2026 preview); VS Code native |
| Entry pricing | Pro $20/mo | ChatGPT Plus $20/mo | Copilot Pro $10/mo + $10 AI Credits |
| Pricing shape | Tiered credits (Pro / Max / Team) | Rolling 5-hr usage windows; tier multipliers | Usage-based credits (June 1 2026); 1 credit = $0.01 |
| June 2026 change | SDK credits separate from chat credits (Jun 15) | Stable as of May 2026 | Pivot to usage-based AI Credits (Jun 1) |
| Standout | Stripe migrated 10k lines in 4 days vs 10 weeks manual | codex-1 RL-trained on real PRs; cloud-parallel | Self-iterating code review before opening PRs |
Below the surface noise, each tool has made one structural commitment that the others can't replicate without redesigning. These are the differences worth weighting.
Claude Code commits to Claude. Codex CLI commits to OpenAI. Copilot CLI deliberately doesn't commit — you pick the model per task via /model. For teams hedging on model choice (or wanting open weights through frontier without rewrites), Copilot CLI's design is the only one that fits structurally. For teams that have already picked their lane, Claude Code or Codex CLI win on first-party depth.
Claude Code and Copilot CLI are local-first — the agent runs in your terminal, edits your files, runs your tests. Codex CLI defaults to cloud-sandboxed execution — the agent works on a copy in OpenAI's sandbox while you keep coding locally. This makes parallel work natural: fire off three tasks in three sandboxes, review the PRs when they're done. For teams who want "kick off a migration and forget about it for an hour," Codex CLI's cloud-parallel pattern is genuinely different.
The Codex CLI binary itself is open source (Rust, npm-installable, Homebrew). Read the source, fork it, swap binaries. The proprietary part is the model behind the OpenAI API it calls. Claude Code and Copilot CLI are fully proprietary CLI binaries that call proprietary models. For teams that want auditable agent-side code (security-sensitive environments, regulated industries), the Codex CLI's OSS posture is unique among the three.
Three real tools, three real fits. The right pick is rarely "which is best overall" — it's "which matches the team you have, the model you're already paying for, and the workflow shape that fits your repo."
codex-1's RL-on-PRs training matters for your use caseAll three are subscription-bundled with chat tools the team probably already pays for. It's common to see Claude Code for deep refactors, Codex CLI for cloud-parallel tasks, and Copilot CLI for GitHub-native PR work — all on the same machine, swapped per task. Cost of running multiple: minimal (each is included in its parent subscription). Cost of context-switching between mental models: real but bounded.
On 19 May 2026 Google announced that Gemini CLI is being transitioned to Antigravity CLI — part of the broader Google Antigravity 2.0 agent-first development platform.
Timeline. Antigravity CLI is available to all users from 19 May 2026. Gemini CLI and the Gemini Code Assist IDE extensions stop serving free and paid individual users (Google AI Pro / Ultra) on 18 June 2026. Enterprise customers on Gemini Code Assist Standard / Enterprise keep Gemini CLI indefinitely, with continued model updates.
What carries over. Agent Skills, Hooks, Subagents, and Extensions (renamed plugins). Antigravity CLI is Go-based (faster execution), adds asynchronous multi-agent workflows, and shares architecture with the desktop Antigravity 2.0 app. Perfect feature parity is not there at launch.
Google's stated reason: individual workflows have outgrown the 2025-era single-agent CLI; users now need multiple agents collaborating. Google announcement →
If subscription-gated CLIs don't fit, the credible OSS alternatives in mid-2026:
— Aider — Python, git-native, terminal-first; brings your own model API key (Claude, GPT, Gemini, Ollama).
— Cline — MIT-licensed VS Code agent; 1.5M+ installs by April 2026; autonomous multi-file execution.
— Gemini CLI → Antigravity CLI — Google's offering, transitioning to Antigravity CLI from 19 May 2026 (see the update note above). Individual users migrate by 18 June 2026; enterprise keeps Gemini CLI.
— Continue.dev — open-source IDE plugin; multi-model.
— OpenHands, Goose, Roo Code, Qwen Code, Kimi CLI — various OSS options.
— Cursor (Cursor 3 released April 2026) — IDE-first with cloud agents, /worktree, parallel Agent Tabs; not CLI-shaped but covers the same workflow.
Cost ceiling matters more in SA than in US-based teams. At ZAR/USD exchange rates and SME margins, the difference between $10/mo and $200/mo per developer is material. Copilot Pro at $10/mo (with $10 AI Credits) is the cheapest entry. Claude Pro and ChatGPT Plus at $20/mo are the next rung. Heavy-usage tiers (Pro+ / Max / Pro200) only make sense once you've proved the workflow pays off.
Local model support matters when bandwidth or POPIA-sensitive data is in scope. None of these three CLIs supports fully-local inference natively — the closest is using them as the framework while routing through a model gateway that hits Ollama / vLLM. For "code never leaves the building" environments, Aider + Ollama is the cleanest path.
GitHub is the dominant SCM in SA dev shops. That alone gives Copilot CLI a default-pick gravity for many teams. The friction of "set up GitHub auth" is zero (already done); the friction of "set up a third subscription" is real.
Talent pool consideration. Engineers in SA in 2026 are most likely to have used Cursor or Copilot before any of the three CLIs. Onboarding cost for Claude Code or Codex CLI is real — not insurmountable, just real. Plan a sprint of pair-programming for whichever tool you pick.
All three move fast — verify versions against the date on this leaf.