OpenAI's open-source coding agent built in Rust. Terminal-Bench 77.3% (#2), SWE-bench Pro standardized 41.04% (GPT-5.2-Codex). GPT-5.4 shipped March 5, 2026. Codex Security agent adds appsec capabilities. 3-4x more token-efficient than Claude Code, 240+ tokens/sec. Free with ChatGPT subscription, sandbox-first execution. 1M+ first-month users. Cleanest security record in Tier 1 — no documented incidents.
Where it wins
Terminal-Bench 77.3% (GPT-5.3-Codex) — strongest on terminal-native tasks (ForgeCode leads at 81.8%)
3-4x more token-efficient than Claude Code (Morph study), API pricing 60-75% cheaper ($1.50 vs $5.00 input)
Sandbox-first execution — genuine safety differentiator (caught SSRF that Claude Code missed per Blake Crosley)
Free with ChatGPT subscription — lowest barrier to entry
619 releases in 10 months — fastest iteration in the category
240+ tokens/sec throughput, 1M+ first-month users (Morphllm)
No documented security incidents — cleanest trust record in Tier 1
GPT-5.4 shipped March 5, 2026 — Codex Security agent adds appsec capabilities
Native macOS + Windows app (Feb-Mar 2026) for parallel agent management
Where to be skeptical
SWE-bench Pro standardized: 41.04% trails Claude Code's 45.89% by ~5pp. First-pass code quality trails Claude Code (67% vs 95% in blind tests)
Users report 20+ minute task times where Claude completes in <1 minute (HN: 104 pts thread). 'Too slow' is the consistent UX complaint
Tied to OpenAI models only
Headline 56.8% SWE-bench Pro number uses custom scaffold — not comparable to standardized results
Editorial verdict
#2 coding CLI. Rust rewrite eliminates Node.js dependency — unique in category. Terminal-Bench 77.3% (#2) and 3-4x more token-efficient. Cleanest security record among Tier 1 tools. GPT-5.4 shipped March 2026. Trails Claude Code by ~5pp on SWE-bench Pro standardized (41.04% vs 45.89%) and first-pass quality (67% vs 95%).
Source
Videos
Reviews, tutorials, and comparisons from the community.
OpenAI Codex CLI
Getting started with Codex
Using OpenAI Codex CLI with GPT-5-Codex
Introducing the Codex app
Related
Coding CLIs / Code Agents
High-volume daily coding, speed-sensitive workflows, air-gapped/locked-down environments (Rust binary)
Software Factories
Developers in the OpenAI ecosystem wanting a fast, token-efficient coding CLI with strong benchmark scores
Claude Code
98Anthropic's official agentic coding CLI. v2.1.81 (Mar 20) shipped `--bare`, smarter worktree resume, and improved MCP OAuth while the repo crossed 82,204 stars and logged ~14 commits/week across 10+ maintainers. Terminal-native, tool-use-driven, with deep file system + shell access, #1 SWE-bench Pro standardized (45.89%), ~4% of GitHub public commits (SemiAnalysis), $2.5B annualized revenue. 8M+ npm weekly downloads. Opus 4.6 with 1M context.
OpenHands
88Category leader in multi-agent orchestration — 69,352 stars (verified), $18.8M Series A, AMD hardware partnership, 455 contributors, 1M downloads/month PyPI (3.4M all-time). SWE-Bench Verified 72% with Claude 4.5 Extended Thinking (updated 2026-03-19), Multi-SWE-Bench #1 across 8 languages. Gap to #2 is enormous on every axis.
OpenCode
88Open-source AI coding agent from SST. v1.2.27 active (2026-03-16) — development resumed after a gap. OpenAI official partnership following the Anthropic OAuth block controversy. 126K+ GitHub stars (star surge driven by Anthropic controversy). Known unauthenticated RCE fixed in v1.1.10+ (CVE, 432 HN pts). CVE-2026-22812 (CVSS 8.8-10.0) is a second serious security incident.
Gemini CLI
88Google's open-source terminal agent with Gemini 3 models, 1M token context, built-in Google Search grounding, and the best free tier in the category (60 req/min, 1K req/day). v0.35.0 (Mar 24) shipped keybinding, policy, and telemetry fixes while the repo hit 98,957 stars and 12,593 forks. Terminal-Bench 2.0: 78.4% (#1). SWE-bench Pro standardized 43.30% (#3). Plan Mode added March 2026. First-pass correctness ~50-60% (Educative.io).
Public evidence
Codex surpassed 1.6M weekly active users (tripled since GPT-5.3 launch). Enterprise adopters include Cisco, Nvidia, Ramp, Rakuten.
90%+ of Codex code generated by Codex itself. Engineers run 4-8 parallel agents. Written in Rust for performance. 1M+ weekly devs, usage up 5x since Jan 2026.
Massive launch discussion. Top comments noted Claude Code outperformed early Codex on docs tasks. Shows improvement trajectory since launch.
Codex CLI leads Terminal-Bench 2.0 (62.9% vs Claude Code's 58.0%). Uses 3x fewer tokens. But Claude leads on OSWorld-Verified for GUI/computer-use tasks. Note: the 77.3% figure belongs to the Droid scaffold, not Codex CLI directly.
Ongoing independent coverage: tested porting JustHTML with Codex CLI, documented skills adoption, GPT-5 model improvements.
Codex CLI scores 77.3% with GPT-5.3-Codex on Terminal-Bench 2.0 — strong but no longer #1. ForgeCode (running Opus 4.6) leads at 81.8%. Codex CLI remains the strongest consumer CLI on terminal-native tasks.
Codex CLI scores 41.04% on SWE-bench Pro standardized, trailing Claude Code's 45.89% by ~5pp. The 56.8% score uses custom scaffolding and is not comparable to standardized results.
Raw GitHub source
Constrained peek so you can sanity-check the source material without leaving the site.
Install globally with your preferred package manager:
# Install using npm
npm install -g @openai/codex
# Install using Homebrew
brew install --cask codex
Then simply run codex to get started.
Each GitHub Release contains many executables, but in practice, you likely want one of these:
codex-aarch64-apple-darwin.tar.gzcodex-x86_64-apple-darwin.tar.gzcodex-x86_64-unknown-linux-musl.tar.gzcodex-aarch64-unknown-linux-musl.tar.gzEach archive contains a single entry with the platform baked into the name (e.g., codex-x86_64-unknown-linux-musl), so you likely want to rename it to codex after extracting it.
Run codex and select Sign in with ChatGPT. We recommend signing into your ChatGPT account to use Codex as part of your Plus, Pro, Business, Edu, or Enterprise plan. Learn more about what's included in your ChatGPT plan.
You can also use Codex with an API key, but this requires additional setup.
This repository is licensed under the Apache-2.0 License.