Claude Code, Codex, Gemini CLI, or OpenClaw: how to pick a 2026 AI coding agent by price, models, Git, and automation

First, align the map: these four are not the same kind of product

Stuffing Claude Code, Codex, Gemini CLI, and OpenClaw onto one “who wins” leaderboard is the most common buying mistake of 2026. The first three are coding agents that work against a repo: they read files, edit files, run commands, and touch Git. OpenClaw is a gateway that works against people and channels: Telegram / Slack / WhatsApp come in, cron and webhooks wake it, and it hands the job to the model or coding agent you configured. The software itself is MIT, with no paid seats. The foundation does not sell you a plan.

Two pieces on this site are easy to mash into this one, so draw the lines now: Cursor vs. Claude Code vs. GitHub Copilot: a year of subscription cost is an IDE-completion stack bill; Hermes vs. OpenClaw is how you install an open-source agent framework. This article answers one question: if you need an agent that writes code, touches Git, and can hang automation in 2026, whose seat do you buy first among these four?

Official doors, in order: Claude Code at code.claude.com/docs; Codex at openai/codex; Gemini’s personal-tier migration note on the Google Developers Blog; OpenClaw at docs.openclaw.ai. Numbers below are what you can check on public pages as of 2026-09-18. Quotas and model names move. Open the pricing page before you buy. Do not treat this table as an invoice.

Price: subscriptions, free quotas, API, and the hidden bill

All four price tags can be written as “about $20.” You are not buying the same thing. Claude Code rides a paid Claude seat and shares quota with web chat. Codex rides a ChatGPT seat or an API key; cloud review and the local CLI may not even share a pool. Gemini CLI used to be nearly free for individuals; that door is closed. OpenClaw’s software is $0. The bill lives entirely in the model and the machine that never sleeps.

Anthropic is blunt: Claude Code is in every paid plan and shares usage with chat. The free tier has no Code. Heavy sessions can move to Console and pay per token so you do not burn a Max seat on one refactor. Codex officially splits into two paths: sign in with ChatGPT and you get the CLI / IDE / web plus some cloud features; use only an API key and you pay per token, but you do not get GitHub auto-review, Slack, or the other cloud pieces. ChatGPT Work and Codex also share quota. Workspace chat is not a second free card.

The Gemini line changed personality once in 2026. Before the cutover, a personal Google login was roughly 1,000 requests a day—that was the acquisition pitch. From June 18, personal Free, Google AI Pro, and Ultra stopped serving requests to Gemini CLI and the Code Assist IDE plugin. Organizations that hold Gemini Code Assist Standard / Enterprise, or that use a paid Gemini / Enterprise Agent Platform API key, are officially unaffected. If you still want a Google agent in the terminal as an individual, the path is Antigravity CLI, and the quota language flipped from “requests per day” to “compute per week.” Do not budget from old blog posts that still say “1,000 free requests/day.”

OpenClaw’s price is the easiest to undercount. Free software is not the same as cheap. You bring your own model: Claude, GPT, or local weights all plug in, but tokens and electricity are real money. The Gateway is meant to stay up; closing a laptop lid is the same as taking every channel offline. If you already pay for Claude Pro or ChatGPT Plus, the marginal software cost really is close to $0—but whether you can legally push subscription tokens through a third-party gateway is a terms-of-service question, not a default “I already paid $20, so I can forward forever.”

For a year-long total, the Cursor / Claude / Copilot piece on this site already does the IDE-stack math. This article only adds one sentence: the real coding-agent bill = the seat + overage tokens + the machine that does not sleep. Compare sticker prices only, and you will pick the combo that looks free and has nobody on call at night.

Models: defaults, context, and whether you can switch

Default model names get a new coat of paint every quarter in 2026. Writing “it must be Opus 4.8” is wasted ink. Shop three things: whether the default is a coding model you actually trust, whether the context window can swallow a mid-size repo in one pass, and whether you can switch models without switching clients.

Claude Code’s edge is not “another 100k tokens.” It is the same engine showing up in the terminal, VS Code, JetBrains, desktop, and the web. A repo-root CLAUDE.md, Skills, and MCP are shared across those surfaces. Project rules do not vanish when you change windows. Codex’s edge is a reviewable client, a default model cut for coding work, and a cheap codex -m switch. The Gemini line’s edge used to be “free plus a million-token context.” After the personal tier shut, that edge is left for people who still hold an enterprise license or a paid key.

OpenClaw is almost designed backwards on models: it does not sell you one. The docs say use the provider you pick. Quality follows the model and the safety policy you are willing to pay for. That is good if you already hold two keys—the same Gateway can cut models by job. It is bad if you want one command and then start typing: there is no default brain. You finish onboarding first.

Model forecasts and the next Gemini number live in a separate piece on the Gemini Pro 4 release window. This article will not bet next quarter’s name. It bets one stabler thing: does the client still live after you change models? Closed suites bet the vendor. OpenClaw bets that you will change the plug.

Agent: permissions, sandboxes, Skills, and multi-agent

“Agent” is a worn-out word in 2026. What decides whether you hand it a repo is the permission model—not the green check in a demo video.

Claude Code is a productized harness: CLAUDE.md as the project constitution, Skills folding repeat work into commands like /review-pr, Hooks running your scripts before or after an edit or commit, MCP talking to Jira / Slack / homegrown tools. Sub-agents can run in parallel; desktop and web can also schedule. Permissions are stricter than a year ago, but they are still “OS rights plus a product confirm,” not Codex’s default no-network sandbox. The official install is curl -fsSL https://claude.ai/install.sh | bash; then claude inside the repo.

Codex CLI’s default stance looks more like security engineering: sandbox policy is read-only / workspace-write / danger-full-access, and the network is denied unless you open it. For daily local use, the official suggestion is --sandbox workspace-write --ask-for-approval on-request. Need extra directories? Use --add-dir, not --yolo as the first move. /permissions shows the current writable roots. MCP exists. Plugins exist. The default trust boundary is still narrower than Claude Code. How much RAM a machine actually spends is a hardware bill we already measured in what projects run on 8GB / 16GB / 32GB—that is not this article’s axis.

# Claude Code
curl -fsSL https://claude.ai/install.sh | bash
cd your-repo && claude

# Codex CLI
curl -fsSL https://chatgpt.com/codex/install.sh | sh
cd your-repo && codex --sandbox workspace-write --ask-for-approval on-request

# Individual Gemini CLI requests have stopped; enterprise / paid keys still work on the original client.
# Personal migration (follow the official announcement):
# curl -fsSL https://antigravity.google/cli/install.sh | bash

# OpenClaw (self-hosted gateway, not a coding REPL)
# Node 24.16+ or 26.x, then openclaw onboard
# Docs: https://docs.openclaw.ai/

Where Gemini CLI still runs, trusted folders plus YOLO mode are the personality: fast to start, boundaries you draw yourself. Hooks, sub-agents, and MCP are all named in the migration note as “things Antigravity is supposed to carry over.” The personal-user question is no longer “is Gemini CLI good?” It is “am I still on an unaffected license?” If you are not, follow the official migration. Do not keep debugging a login that already stopped serving requests.

OpenClaw’s agent work sits on another layer: sessions, memory, multi-agent routing, skills, and channel plugins. It can call a coding agent. It can also call a browser, Gmail, or GitHub. The security model is “trusted gateway, untrusted execution, write the policy down hard”—which means you review ~/.openclaw/openclaw.json, the channel allowlist, and webhook tokens yourself. Community write-ups that say “do a security review before you deploy” are not being polite. An unrestricted Gateway on the public internet is handing model keys and repo operations to anyone who can hit the port.

Git: commits, PRs, and CI—who actually fits

Whether a coding agent is worth paying for lands on Git: will it write a commit message, open a PR, and finish a headless review in CI? “Look at this snippet” in a chat window is not Git integration.

Claude Code treats Git as a first-class citizen. The official overview literally demos claude "commit my changes with a descriptive message" and piping git diff into a headless review. The CI story is “an event comes in, the agent opens a PR or leaves a comment.” Codex splits local and cloud: local codex is for watching the diff; codex exec is for scripts; GitHub auto-review is a cloud feature, and OpenAI writes that only reviews run through GitHub count against Code Review quota. API-key-only users should not expect that layer to appear for free.

The Gemini Git story now forks. Enterprise Code Assist and the GitHub products on Google Cloud are still there. Personal terminal users have to carry the old GEMINI.md, MCP, and skills over to Antigravity. The migration note promises to copy config; it does not promise 1:1 feature parity—Google does not claim that itself. Teams that wired a critical release pipeline to personal Gemini CLI should already have cut over after June. If you have not, this is priority work, not an optional polish.

OpenClaw touches Git unlike the other three. It is not “open a TUI inside the repo.” It is someone in a channel saying “draft a PR for this issue,” and the Gateway using a configured gh and workspace to do it. That fits on-call, a product manager dropping a bug in Slack, or a midnight webhook waking a hotfix. It does not fit you sitting in the repo for a three-hour refactor—that job belongs to Claude Code or Codex. OpenClaw’s job is to send the result back to the channel.

Automation: headless, cron, channels, and always-on

Automation is the axis that actually opens a gap. All four can look competent while you are watching. Only some combinations keep working after you close the lid.

Claude Code’s automation has three layers. In the terminal, claude -p follows Unix: pipe logs, batch-translate, review changed files in GitHub Actions. Desktop and web have Routines; cloud schedules can still run after the computer is off, and they can hang off GitHub events. Channels (Slack / Discord / web hooks) can push an external event into a session. That is vendor-hosted “it keeps going after you leave your chair”—if you can afford the seat, and if you accept the job running on their cloud.

Codex automation leans toward engineering scripts: codex exec is headless, can emit JSON, and fits a runner you own. Cloud tasks can be submitted from the terminal and pulled back locally. GitHub and Slack are cloud capabilities of a ChatGPT seat, not a gift on the API-key path. It is a good fit for “this pipeline must be reproducible and auditable.” It is a poor fit for “I live with an assistant on WhatsApp.”

Gemini / Antigravity automation depends on which license you still hold. The enterprise side keeps going through Google Cloud and Code Assist. The personal side has to redesign jobs like “scan the repo once a day” against the new CLI’s weekly compute cap. The old “1,000 requests a day is enough for cron” spreadsheet is dead.

OpenClaw was almost born for this axis. The docs split automation into cron, heartbeat, hooks, webhooks, Task Flow, and standing orders. Cron can send results back to a channel or an HTTP endpoint. Heartbeat, by default, sweeps the inbox and the checklist about every thirty minutes. Gmail Pub/Sub and the Gateway’s /hooks/wake exist so an outside system can wake the agent. The channel list is long: Discord, Slack, Telegram, WhatsApp, iMessage, Matrix, Teams… the chat tool you already use often has a plugin. How to keep a Gateway standing on a remote Mac is a separate hands-on: OpenClaw on a remote Mac, US East and West. This piece will not repeat the install steps.

Always-on automation has a physical premise: the process cannot go to sleep with the laptop lid. Claude / Codex cloud schedules hand that premise to the vendor. OpenClaw hands it to your machine. A dedicated Apple Silicon box you can SSH into, with keys living in the keychain, is more honest than opening four agents on the thin laptop you take through airport security.

How to choose: one decision table, one install or two

Do not start from a review-site total score. Start from what will actually happen this week.

One person, one main repo, a few hours a day: pick Claude Code or Codex, not both. Already paying for Claude Pro? Take the former. Already paying for ChatGPT Plus, and you care about the sandbox and an open-source client? Take the latter. Do not stack two paid seats “so you do not miss out.” That is how a hidden bill becomes a subscription.

A small team with on-call and scheduled jobs: put the coding agent on a local machine or in GitHub Actions, and put OpenClaw on a machine that does not sleep. The Gateway only wakes, only routes, only replies in the channel. The right to actually edit code stays inside the coding agent’s sandbox or confirm flow. If a self-evolving framework like Hermes is also on the shortlist, read the Hermes vs. OpenClaw piece on this site first and decide the framework layer there. Do not mash it into this article’s “terminal coding seat” purchase.

Enterprise buying should not follow a personal blog’s “free or $20.” Seats, data residency, whether a third-party gateway may use subscription tokens, and which identity the CI bot uses to push a branch decide whether you can ship—before the model name does. Price tables get a new edition every year. Permissions and Git identity hurt more to change.