Codex CLI
Open source, about 106,000 stars; a simple task drops to a cheaper tier without a human decision.
Verified: Sep 29last change: Sep 28next price check: Dec 28
Facts with sources
- One agent across terminal, IDE, web, and cloud. Source (in Russian)
- Terminal-Bench 4.0: Codex + GPT-6 Astra on high effort solves 57.88% of tasks at $2.3K per run; switching to max adds 0.3 points and a thousand dollars. Source (in Russian)
- OpenAI blocks access by IP; most popular VPNs are detected. Source (in Russian)
Strengths
- Across Tech With Tim's five tasks Sol matched Opus 5 on correctness but ran 4 to 7x faster; the bug fix was two lines.
- The $20 Plus plan includes the flagship GPT-5.6 Sol; in Tim's experience the same task burns 3 to 4x less quota than in Claude Code.
- Open source under Apache 2.0, with sandboxing, AGENTS.md, subagents, MCP and a non-interactive codex review, per OpenAI's docs.
- In Soham Kamani's demo three subagents took a Jira ticket to a GitHub PR; the reviewer agent caught an API regression before the PR.
- Everything runs in a sandbox on your machine by default and any command outside it asks for permission; cloud runs use an isolated VM, the same one ChatGPT Work gets.
Weaknesses
- Does exactly what was asked: on the habit tracker it wrote far fewer tests and checks than Opus 5 with its 27 tests.
- Cloud only: no offline mode, and speed depends on the network and OpenAI's queue (Alex To Go Eng).
- Limits keep moving: in July OpenAI temporarily dropped the 5-hour cap, leaving a weekly one; Plus gets 10 to 100 Sol messages per 5 hours.
- The agent edits files itself: the reviewer advises a dedicated folder and a least-privilege key so it cannot touch production data.
- The Codex lead says the harness is "always a step ahead of the model" and shrinks over time; a feature you buy today can vanish after the next model, as already happened with /goal.
Example from practice
From a Jira ticket to a GitHub PR with no human in the loop
Soham Kamani wired Jira and GitHub into Codex through MCP servers and defined three subagents: one reads the ticket, one fixes the code, one reviews. On the ticket 'discounts increase the price instead of reducing it' the agent found the sign error, the reviewer bounced the first fix over an API regression, and then a PR was opened. The agent then posted the PR link into the ticket and moved it to In Progress. The video gives no timing or cost.
source: Soham Kamani, "Going from JIRA Ticket → GitHub PR using MCPs (with OpenAIs Codex)"
What reviewers say
Reviewer opinions on their own tasks and hardware, not measurements. Links go to the videos; transcripts live in docs/raw-data.
- Codex vs Claude – an Honest Comparison · Tech With Tim, Aug 27The reviewer's take: choose Codex for speed and quota efficiency, since it matched Claude on correctness, ran faster and offers the flagship model at $20.
- Going from JIRA Ticket → GitHub PR using MCPs (with OpenAIs Codex) · Soham Kamani, Aug 26The author shows that with Jira and GitHub MCP servers and three subagents Codex can carry a ticket through to a PR on its own; a demo, not a comparison.
- OpenAI Codex CLI: Terminal-Based Coding Agent Setup and Workflow · Alex To Go Eng, Sep 5The author sees Codex CLI as the cure for copy-pasting code into a chat window, but insists on an isolated folder, least-privilege keys and mandatory human review.
- Codex vs Claude Code vs Cursor vs Copilot – Best AI Coding Agent in 2026? · Byte & Beyond, Aug 31Without running any code, the author ranks Codex first for automation because parallel agents, worktrees and scheduled background tasks are core to the product.
- Building Codex with Tibo Sottiaux · The Pragmatic Engineer, Sep 9Tibo Sottiaux (Codex lead) on The Pragmatic Engineer: the harness is open and not tied to OpenAI models, inside OpenAI the security review blocks PR merges automatically, and the human role in review shifts to discussing intent; a vendor on its own product.
Who it is for
- Engineering team that cares about open source
- Team already on a ChatGPT plan with billing in place
Models in this agent
Score is from our manager-task benchmark; models without a score are not measured yet.
What changed
- OpenAI cancelled the release of Astra 6.1 after it failed internal safety testing →
- OpenAI is preparing a $500-a-month ChatGPT Pro Max plan, according to TestingCatalog; no official announcement yet →
- OpenAI released GPT-6 Sol and GPT-6 Luna: API prices half of the GPT-5.6 promo rates (Sol at $2 and $10 per million tokens) →
- Codex CLI 0.157.0 added GPT-6 Sol and GPT-6 Luna, including Amazon Bedrock support and migration prompts for older models →
- OpenAI paused sales of the $200 Pro subscription because of Astra demand; no new Codex seats on that tier for now →
- Codex lead: ChatGPT Work is the full Codex harness in a cloud VM with internet access; the harness is open and not tied to OpenAI models →
- OpenAI released GPT-6 Astra: the top tier in Codex, 2.5 times the price of GPT-5.6 Sol →
How to brief an agent like this
The four-line brief is the Foundation module: prompt structure, LLM agents and MCP, an AI policy for the team. It starts with the free open module.