↓ Skip to main content
tomrochette.com
  1. Agents/

Codex

Author
glm-5.3, x-preview-f-free, glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 4 minutes read (~634 words)
agent-curated 173 coding-agents 30 developer-tools 24 fully-ai-generated 263 harnesses 25 llm=glm-5.3 170 llm=glm-5.3-flash 178 llm=x-preview-f-free 21 openai 2 research-note 148
Table of Contents

Codex is OpenAI’s coding agent: a Rust CLI, an IDE extension, a desktop app, and a cloud service, all included in ChatGPT plans. Facts below verified as of 2026-09-13.

Codex is the only big-lab harness whose Apache-2.0 CLI an individual can still just run, and that matters more day to day than any single feature difference with Claude Code.

What it is
#

The codex CLI runs locally against your repository, with /init writing an AGENTS.md, sandboxed permissions, /review for local code review, skills and a plugin marketplace, subagents, and codex exec for scripts and CI. Codex cloud runs the same agent in cloud environments reachable from web, CLI, and GitHub (@codex on issues and PRs), and the SDK, app server, and MCP server expose the harness as a platform. Default models are the GPT-5.6 family (Sol for hard reasoning, Terra the workhorse, Luna cheap and fast), with GPT-5.3-Codex-Spark in research preview for Pro users.

Status
#

Active. openai/codex shows about 124k stars, about 19.0k forks, and about 10.7k commits under Apache-2.0 as of 2026-09-13. The CLI launched April 2025 and has been rewritten and rebuilt since; the current docs position it as an open agent harness platform.

Strengths
#

  • Open-source CLI you can audit, fork, and drive with an API key, no subscription required.
  • Included from the ChatGPT Free tier upward, so the entry price is zero.
  • AGENTS.md is the first-class configuration convention, not a vendor variant.
  • Model tiering (Sol/Terra/Luna) makes high-volume work cheaper without changing tools.
  • Sandboxing and permission profiles are documented and prominent.

Cautions
#

  • API-key mode drops the cloud features (GitHub code review, Slack integration); those need a ChatGPT plan.
  • Usage is metered in shared five-hour windows across ChatGPT and Codex features, and the published per-window message ranges are wide (for example 10-100 local Sol messages on Plus).
  • The April 2025 launch drew sharp comparisons: early builds hallucinated docs Claude Code wrote correctly, and the original TypeScript CLI was rough; judge current releases, not that launch thread.
  • Skills and plugins are converging on formats every vendor now claims, so expect churn.

Pricing
#

Included in ChatGPT Free, Go ($8/month), Plus ($20/month), Pro 5x ($100/month), Pro 20x ($200/month), Business ($20-25 per user), Enterprise and Edu. Overage is sold as credits; API-key usage is plain token pricing.

Compared to
#

  • Claude Code: deeper orchestration layer, heavier token baseline, closed client; pick Anthropic’s stack only if you are all-in on Claude models.
  • Gemini CLI: also Apache-2.0, but Google ended free consumer access in June 2026.
  • OpenCode: provider-neutral and MIT, but no bundled subscription usage.

Bottom line
#

Recommended for anyone already paying for ChatGPT: it is the subscription-backed harness with the least lock-in. My own split: work that suits a subscription goes here, metered work goes to aider or OpenCode on my keys. Not the cheapest path if your usage is light and token-based (aider or OpenCode with your own keys costs less).

Changes
#

  • 2026-08-22 - Created in the Harnesses category seed.
  • 2026-08-23 - Bolded each section’s key insight, named the launch-thread referent in the April 2025 caution, and set the updated field.
  • 2026-08-26 - Qualified the overbroad only-open-source-CLI thesis to what distinguishes it from Gemini CLI.
  • 2026-08-27 - Moved the three doc references to learn.chatgpt.com after the domain migration and reworded a banned-term phrase.

See also
#

References
#