Codex
OpenAI AI agent: takes tasks in batches and handles them in parallel.
Sign in $0
Included in ChatGPT plans
GPT models
OpenAI models only
Not directly from Russia
Requires a foreign IP address and card
Code partly open
CLI is open, models are not
Use it to hand off many tasks at once — sign in with a regular ChatGPT account, even a free one. Do not use it for editor suggestions and polished screens: those are the domains of Cursor and Claude Code.
Codex is not a chat about code, but an autonomous executor. You describe the task in words, then the agent works on its own: finds the right place in the project, makes changes, checks its work, and brings you a finished result for review. Its main advantage is that you can launch several tasks at once. Each works in its own copy of the project and does not interfere with the others. It was born in the terminal, and since July has lived right in the ChatGPT app, alongside regular chat.
Where Codex sits on the big AI map
Nearby: Claude Code·Cursor·GitHub Copilot·Lovable·Replit·+ more 564
1. Should you get it
The choice comes down to two profiles: see which one is yours.
Who it is for. Codex is built for a flow of work: many tasks launched at once and almost for free. Hand it a dozen small fixes in the morning, clear out stale changes, review someone else's work — the agent handles several tasks at once, each in its own copy of the project. Sign in with a regular ChatGPT account, even a free one. You can assign tasks from the app, browser, and phone.
Don't write code? A terminal is not required: assign a task in the ChatGPT app or browser, just like in chat. At OpenAI, they themselves call the agent a superpower for people without development experience
Who it is not for. There are no suggestions while you type code — go to Copilot and Cursor for that. OpenAI models only; you cannot connect your own local ones. Claude Code is where people go for design and polished screens: AI's weakness in design
Just getting started? Haven't written code or built projects yet - start with the page“What is vibecoding”, then come back here for a tool.
If it fits, install it using section 3. If you are choosing between it and Claude Code, see section 7.
Codex News
- Capy speeds up PR work: Garry Tan estimates at least a 4x gainSeptember 23
- AI Agent Reached the Wall: Codex Helps Build a Voice Assistant on Raspberry PiSeptember 23
2. What it does
The easiest way to show it is in practice: in our project, Codex is the night shift. Below are six real tasks it completed here over the last month.
Watches the sources. Several times a day, the agent rereads 15–33 pages where AI news emerges and compares them with yesterday's. On August 17, it went through 15 sources and picked nothing — it was all noise. This refusal is work too: junk does not make it into the feed.
Writes entire news stories. The agent put together a story about Lovable starting to deliver apps in ChatGPT itself: a verdict, what changed for the reader, and sources.
Judges every story. A judge stands before the writer, and that is Codex too: it scores every story, separates urgent from ordinary, and decides what is not worthy of the feed.
Draws covers at night. The agent generated four clay-style course covers overnight, all four in a consistent style on the first try. The pipeline orders news covers from it on its own, without a person.
Clicks through panels like a person. When the site's hosting needed a different plan, the agent opened a browser, entered the control panel, and switched settings with buttons. It checked afterward: the configuration was applied.
Fixes and checks code. Dependency updates for security, fixing a broken chart on the ratings page, a new screen in the internal dashboard — the agent handles routine engineering work on its own.
What it looks like. The cycle is simple: you assign a task in words from the ChatGPT app, terminal, or phone — the agent edits files in a sandbox with networking disabled, shows every change line by line, and takes it to a finished pull request. Several such tasks can run at once: in the cloud, the agent makes parallel attempts and shows the best one.
How much freedom to give it. By default, the agent asks for confirmation on risky steps; you choose the boundaries from there — up to a complete “do not ask.” The agent knows the line too: in a case shown by Greg Brockman, Codex assembled and launched an ad campaign, but stopped at the payment button and asked for permission.
Then on its own. Higher up the axis is goal mode: put a goal in front of the agent, not a list of steps. Plus computer use: Codex clicks through your apps on Mac and Windows — writes throwaway scripts instead of live clicks, which is faster. Using this setup, GPT-5.6 Sol in Codex in five hours completed the daily challenge itself in Slay the Spire 2. The top level is the cloud: the pipeline works while your laptop is closed, and you can manage it from your phone. How to hand work back to the agent for another pass — loops.
3. Where it works and how to install
One ChatGPT account opens 7+ surfaces: app, terminal, IDE, cloud, GitHub, phone, SDK. Start with the app — since July, Codex has lived right inside ChatGPT on macOS and Windows, and since August 11 on Linux too. The terminal is for people who live in the CLI: it has open source and full control.
Before you start. You need a ChatGPT account — even a free one is enough — or an OpenAI API key. Documentation for all surfaces is on the official Codex page.
You can install it from Russia. First, switch your Russian IP address to a foreign one: Russia is not on OpenAI's list of supported countries, and the service will not open without changing your address. Next, you need a foreign card for paid plans; working payment routes are collected in section 5. Keep the risk in mind: access from an unsupported country may cost you your account. The closest local equivalent by genre is Yandex SourceCraft: agent mode from task to pull request, with payment in rubles.
In the ChatGPT app+
There is no separate Codex app anymore: since July, the agent has lived as a tab in the ChatGPT app — agent view with parallel tasks, schedules, and computer use. macOS and Windows; Linux — preview since August 11.
Download the ChatGPT app at openai.com — the Codex tab is inside; Linux is installed using .deb and .rpm packages from the OpenAI repository.
In the terminal+
An open-source terminal agent: Apache-2.0, Rust, around 106 thousand stars. Named sessions, portable plugins, the latest MCP protocol; sign in with a ChatGPT account.
npm install -g @openai/codexStep 1. Install the package with the command above — or via Homebrew: brew install --cask codex.
Step 2. Run the codex command and sign in with your ChatGPT account.
In the editor+
A VS Code extension — also works in Cursor and Windsurf — plus JetBrains and native support in Xcode. Since July, JetBrains AI has called Codex its recommended default agent.
code --install-extension openai.codexIn JetBrains — enable Codex in the JetBrains AI settings.
In the cloud+
Background tasks on chatgpt.com/codex: nothing to install, code runs in a cloud environment, and the result is a ready pull request. It can make several attempts in parallel and run scheduled tasks.
Open chatgpt.com/codex and connect a repository from GitHub.
On GitHub+
Auto-review of new pull requests, @codex review, and fixes via @codex fix; since late July — diff reviews across multiple repositories at once. codex-action runs the agent in CI.
Enable Codex for a repository at chatgpt.com/codex — then call @codex directly in pull requests.
From your phone+
The Codex tab in ChatGPT on iOS and Android: start tasks, approve changes, and view diffs from your phone — even control your Mac with its screen off.
Install ChatGPT from the App Store or Google Play — the Codex tab is inside.
For your own agents+
TypeScript and Python SDKs plus an app server over JSON-RPC — embed Codex into your products: the VS Code extension runs on this same layer.
SDK and app server documentation — learn.chatgpt.com/docs/codex-sdk.
4. What the tests show
An agent score is always a pair of “agent + model”: Codex is measured with GPT-5.6 and GPT-5.5, Claude Code with Fable and Opus. Compare pairs, not brands.
The best Codex pair is with GPT-6 Astra. In the 23.09.2026 measurement, it scored 61.6 aa index - rank #2 out of 6 pairs in the table. The current leader is Claude Code with Claude Fable 5.1 max (with fallback): 62.2.
| # | Agent | AA score | RuBench | |
|---|---|---|---|---|
| 1 | CodexOpenAI | 253 | 80GPT-5.6 Sol | 66,7GPT-5.5 · xhigh |
| 2 |  Claude CodeAnthropic | 208 | 77Fable 5 | 78,7Opus 4.8 · max |
| 3 |  CursorAnysphere | 187 | 62GPT-5.5 | — |
| 4 |  GitHub CopilotGitHub | 30 | — | — |
| 5 |  AntigravityGoogle | 27 | — | — |
| 6 |  ReplitReplit | 26 | — | — |
| 7 |  LovableLovable | 23 | — | — |
agent table: weight - editorial · AA score - best pair · AA Coding Agent Index 23.09.2026 · Terminal-Bench 4.0 23.09.2026 · DeepSWE v1.1 22.09.2026 · “—” - pair not measured
External measurements show parity. At the top of Artificial Analysis's Coding Agent Index, there is a tie: the best Codex and Claude Code pairs both score 67 (v1.3 scale, August snapshot). On the public Terminal-Bench 2.1, the Codex + GPT-5.5 pair holds second place at 83.1% — first is Claude Code with Fable 5 at 83.8%.
Scale is growing fast. On August 13, OpenAI marked the milestone 15 million users Codex — and reset limits for everyone at once to celebrate.
Choosing between it and Claude Code — the duel in section 7.
5. How much it costs and how to pay from Russia
Access is free. Codex is included in every ChatGPT plan, even Free: create an account and the agent is already there, with minimal limits. There is no separate “Codex” subscription: you pay for the full ChatGPT plan, which also unlocks chat.
Plans and what they include
| Plan | Monthly price | What it gives you |
|---|---|---|
| Free | $0 | Free access to Codex, minimal limits |
| Go | $8 | More limits — the cheapest paid plan |
| Plus | $20 | 10–100 GPT-5.6 Sol messages per 5-hour window — a practical baseline for solo work |
| Pro | $100 and $200 | 5x and 20x tiers — long sessions and unlimited parallelism |
| API | By tokens | GPT-5.6 Sol — $5/$30 per million input/output tokens; Luna after the July cut — $0.20/$1.20 |
prices checked on August 17, 2026 against OpenAI's official pricing page
Limits are shared across everything. The plan shares one pool between ChatGPT chat and Codex: messages in 5-hour windows plus weekly bars. Above the cap, you can buy credits at around $0.04 — usage is billed as a stream, effectively at API prices. For scale: a large prototype build by Simon Willison burned through $23 in tokens at API prices — but ran inside a regular monthly subscription.
Prices are going down. On July 30, OpenAI cut API prices lightweight models: Luna got 80% cheaper, Terra 20% cheaper — the same plan now covers more runs. And in July, limits were reset three times for all paid plans at no charge.
You can pay from Russia. A Russian card will not work, and Russia itself is not on the list of supported countries — bypassing access may cost you your account, so weigh the risk in advance. The working routes are the same as for other services: a foreign card (Armenia, Kazakhstan, Kyrgyzstan), an intermediary for rubles, and a virtual card topped up in USDT. As long as the Free plan is enough, you do not need to pay at all. A step-by-step breakdown of the options — how to pay for AI services from Russia.
What we checked and how
| Fact | What we know | Verified |
|---|---|---|
| Plans | Codex is included in all ChatGPT plans, with no separate subscription. Pro has two tiers: 5x for $100/month and 20x for $200/month. Business - $25 per seat monthly or $20 with annual billing (from 2 seats); Enterprise/Edu through sales. | 2026-08-17 |
| Free access | Yes: Codex is included in the ChatGPT Free plan ($0) with minimal limits. A key difference from Claude Code, which has no free entry point. | 2026-08-17 |
| Limits | The window is shared between local messages and cloud tasks; the pool is shared with ChatGPT chat, with weekly limits also applying. Pro gives ×5 ($100 tier) and ×20 ($200) over Plus. Above the cap - credits (~$0.04 per credit, charged per token - effectively the API price). Issued limit resets are valid for 30 days. On 12.07, OpenAI temporarily removed the 5-hour window for paid plans; the window is still described in the pricing docs. | 2026-08-17 |
| API price | Input/output per 1M tokens; cached input is 10 times cheaper than regular input. On 30.07, OpenAI cut prices: Luna by 80%, Terra by 20%; Sol remained $5/$0.50/$30 - the same as GPT-5.5. | 2026-08-17 |
| Availability in Russia | Russia is not on OpenAI's official list of supported countries (Belarus is not either). Kazakhstan, Armenia, Georgia, and Kyrgyzstan are listed; Ukraine has caveats. Status unchanged since the previous snapshot (10.06). | 2026-08-17 |
All vendors rebuilt billing around token credits in 2026; figures change every few weeks. Each line above has a verification date: do not trust figures without a date, including ours - and check them against OpenAI's official pricing.
Working with AI agents? Here is what it looks like over the long haul63 days, 4,642 commits, not a line by hand. A live feed of this site being built.Watch live
6. Can you trust it with code
What to check before letting an agent into a production repo.
Where the code goes. On personal plans - Free, Plus, Pro - OpenAI may train models on your content by default: disable training with the toggle in Settings → Data Controls. Business plans and the API do not train by default, while Enterprise adds zero data retention. Turned off the main toggle? Check Codex's own settings too: third-party analyses have found a separate control there.
What the agent can break. It edits files and runs commands - but by default in a sandbox with networking disabled: Seatbelt on macOS, bwrap on Linux, and a system sandbox or WSL2 on Windows. Save "no sandbox" mode for extreme cases: in July, it was in this mode that the agent deleted a user's files while overriding the home directory - the bug was acknowledged and fixed.
The agent checks its own code. The CLI is open source - you can read the code to see exactly what the agent does. And since August, Codex has reviewed your pull requests for vulnerabilities too: Security Review in preview, it comments only on P0 and P1 risks - in an independent test, it caught 16 of 23 known vulnerabilities.
What we checked and how
| Fact | What we know | Verified |
|---|---|---|
| Code privacy | On Free, Plus, and Pro, content may be used for model training until the Settings > Data Controls > “Improve the model for everyone” toggle is turned off. Business/Enterprise/API do not train by default; Enterprise supports zero data retention with OpenAI approval. Third-party analyses say Codex has a separate training control for “full environments” - check Codex's own settings. | 2026-08-17 |
| Sandbox | Per-OS sandbox: macOS Seatbelt (out of the box), Linux and WSL2 - bubblewrap, Windows - native Windows Sandbox or WSL2. Network is off by default; access modes: read-only / workspace-write / danger-full-access. The 16.07 case (agent deleted files in sandbox-free mode) is an argument not to disable it unnecessarily. | 2026-08-17 |
| Open source | CLI is open: Apache-2.0, Rust, ~106.4k stars (measured 17.08). Stable version 0.147.0 (07.08.2026), alphas ship almost daily. Install: npm i -g @openai/codex or brew install --cask codex. | 2026-08-17 |
| Local alternatives | There is no direct Russian equivalent of a multi-surface AI agent. The closest local option - SourceCraft Code Assistant (Yandex): agent mode “from task to pull request,” Free plan + Pro 250 rubles/month. | 2026-08-17 |
7. Codex or Claude Code
The fork is not “who is better,” but “what kind of work.” Codex wins on breadth and price: a pipeline you can run in batches, with a regular ChatGPT account as the entry point. Claude Code wins on depth: one agent stays with a complex task in a large repo for a long time and makes careful edits.
In August, both entered the big leagues. IBM is moving to Codex a platform for 150,000 consultants. Samsung, meanwhile, has brought Claude Code into chip development: verification shrank from a month to two days. And on August 11, Codex itself learned to import settings and chats from Claude Code - moving over is now an official button.
What to choose for the task
| Criterion | Codex | Claude Code |
|---|---|---|
| Strength | mass parallel pipeline: cheap, fast, and not just code | complex tasks in a large codebase, meticulous edits |
| Fresh proof | IBM moves 150,000 consultants to development via Codex (13.08) | Samsung: chip verification in 2 days instead of a month (13.08) |
| Entry and price | free entry through ChatGPT, paid plans raise limits | from $17/mo, no free plan |
| Tests | Coding Agent Index 67 (tie at the top); Terminal-Bench 83.1%, #2 | Coding Agent Index 67 (tie at the top); Terminal-Bench 83.8%, #1 |
| Models | OpenAI only (GPT-5.6 Sol is the main model) | Claude only (API/Bedrock/Vertex/Foundry) |
| Openness | open-source CLI (Apache-2.0, ~106k stars) | no, proprietary CLI |
Live install count. For a developer, npm works like an app store: terminal agents are installed from there, and every installation appears in a public counter. The chart adds up installs for the week - the higher the line, the more often it is being installed now.
Installation race: Codex vs. Claude Code
npm downloads per day · ● live
The duels are real, the score is uneven. A chess engine written with Fable 5 beat the Codex engine by 10:0 with the same prompt. A week later, Simon Willison's Codex built a game noticeably better than the Claude version - and then failed to notice a visual bug in it even in screenshots.
Pros keep both. Gergely Orosz, author of The Pragmatic Engineer newsletter, on August 6 gave joint first place in the terminal to both - Codex and Claude Code. The tone shifted, though: DoltHub named Codex its primary agent in its annual review - "five times more work per unit of time" - while design and frontend work still goes to Claude. The norm remains: keep both and split tasks by type.
8. Frequently asked questions
Codex or Claude Code - which to choose?+
Measurements are tied: in Coding Agent Index (Artificial Analysis, v1.3), the best pairs from both score 67. Since July, practitioners have noticeably shifted toward Codex - faster and cheaper for high-volume tasks; design, frontend, and complex architecture still go to Claude Code. The working pattern is to keep both and split tasks by type.
How much does Codex cost?+
Entry is free: Codex is included in all ChatGPT plans - Free $0, Go $8/month, Plus $20/month, Pro $100–200/month. Limits - messages per 5-hour window (Plus on GPT-5.6 Sol: 10–100), above that - credits at roughly $0.04. Via API - metering: GPT-5.6 Sol costs $5/$30 per 1M input/output tokens. Verified 17.08.2026 against the official pricing page.
Does Codex work in Russia?+
Russia is not on OpenAI's official list of supported countries (verified 17.08.2026); access from an unsupported country may lead to an account ban. Working options - a foreign IP address and an overseas card for paid plans; among local options, the closest is Yandex's SourceCraft Code Assistant (agent mode, payment in rubles).
Can you use Codex for free?+
Yes: Codex is included in the ChatGPT Free plan ($0) with minimal limits. For regular work, Plus $20/month is more practical (on GPT-5.6 Sol — 10–100 messages per 5-hour window); for heavy users — Pro $100–200/month. Verified against official pricing on 17.08.2026.
Useful reading before you start
- What is vibecoding: starting from scratch.
- Vibecoding 101: your first agent project - built with Codex.
- Loops: how to give an agent recurring work.
- MCP: how an agent connects to your services.
- How to pay for AI services from Russia.
- The full AI agent ranking.