OpenAI Codex, 2026
Codex Isn't What It Was in 2021.
Today's Codex is OpenAI's agentic coding product: a CLI, IDE extensions and a cloud agent that edit real files and run real tests — bundled into ChatGPT plans including Plus, metered by token credits on a rolling 5-hour window. This guide covers what's included, where the limits bite, and what handles the overflow.
Honest scope
Codex is OpenAI's product — here's our actual place next to it
VUST doesn't resell or wrap Codex, and @vustbot cannot touch your repository. What it does well is the surrounding work: explain the stack trace, compare two architectures, draft the PR description, sanity-check the migration plan — frontier models per-use in Telegram, free tier first, no card. If you want a full agentic-coding switch instead, the /codex-alternative page compares real options.
See the difference
The disambiguation, the plan math, and the mid-task limit — honestly.
Who this guide is for
- ChatGPT subscribers
- Finding out what your Plus/Pro seat already includes
- Codex access is bundled — CLI, IDE extensions and the cloud agent draw from your plan's token-credit allowance. No second subscription.
- Name-collision victims
- Googling 'Codex' and landing in 2021
- The old completion model is retired; today's Codex is an agentic product that edits files and runs tests. This page covers the current one only.
- Limit planners
- Budgeting real agentic work
- Token credits on a rolling 5-hour window are the constraint that matters; heavy users commonly add $100–200/mo in overflow credits.
Codex in three facts
- 01
What it is
OpenAI's agentic coding product: reads your repo, edits real files, runs commands and tests. Surfaces: CLI, VS Code/Cursor/Windsurf extensions, cloud agent at chatgpt.com/codex, desktop and iOS.
- 02
How it's priced
Bundled into ChatGPT plans (Plus $20/mo included; Pro tiers carry more capacity). Since April 2026, usage meters token credits on a rolling 5-hour window; Business/Enterprise can add usage-billed Codex-only seats (June 2026).
- 03
What powers it
GPT-5-series models; the exact default and the selectable variants move over time — the CLI's /model command shows your account's current options.
For the questions that don't need your repo
Open @vustbot for frontier answers per-use — the GPT-5.6 family, Claude Opus 4.8, Gemini — free tier, no card, while Codex's window rolls.
Honest scope — Codex and us
We don't run Codex
No reselling, no wrapping, and @vustbot cannot touch your repository. That honest limit is also why it needs zero setup.
The overflow slot is real
Mid-task credit exhaustion sends the non-repo questions somewhere: explain the trace, weigh two designs, draft the plan. Frontier chat per-use covers exactly that, cardlessly.
Verify before buying
Plan bundling and metering described here reflect OpenAI's 2026 state at build time — check openai.com before purchase decisions; vendors move.
Frequently asked questions
Is Codex included in ChatGPT Plus?
Yes — Codex access is bundled into ChatGPT plans including Plus ($20/mo), with usage drawn from token-credit allowances rather than a separate subscription. Higher tiers (Pro) carry larger allowances, and Business/Enterprise workspaces can add pay-as-you-go Codex-only seats (announced June 24, 2026).
Is this the same Codex as the old code-completion model?
No. The 2021 Codex was a completion model (an ancestor of early Copilot) and is long retired. The 2026 Codex is an agentic product: it plans multi-step coding tasks, edits files, runs tests and iterates — through a CLI, IDE extensions and a cloud agent. Same name, different thing; this page covers the current one.
What models power Codex?
GPT-5-series models, including Codex-specialized variants. OpenAI evolves the lineup continuously, so treat the exact list as a moving target and check the /model command in the CLI for what your account offers — VUST does not restate a version it cannot source.
What are the real usage limits?
Since April 2026 Codex meters token credits on a rolling 5-hour window: heavy agentic runs (large repos, long iterations) consume credits fast, and when the window's allowance is spent, you wait or pay for additional credits. This is the constraint practitioners actually plan around — not a hard monthly cap.
Codex vs Claude Code vs Cursor — which one?
All three are agentic coding tools with different centers of gravity: Codex leans on your ChatGPT plan and OpenAI models; Claude Code is Anthropic's CLI-first agent tied to Claude plans; Cursor is an IDE with agent features and its own credit model. If you already pay for ChatGPT, Codex is the zero-extra-signup option. Our guides on Claude Code and Cursor cover their plan math.
What's the cardless option for the non-repo questions?
Codex needs an OpenAI account and a card-backed plan for meaningful use. For the surrounding questions — explain an error, weigh two designs, draft docs — @vustbot answers with frontier models (GPT-5.6 family, Claude Opus 4.8, Gemini and more) inside Telegram: free tier, per-use Balance, no card. It doesn't touch your repo; that honest limit is also why it needs no setup.
Ready when you are
Codex for the repo. A chat for everything around it.
The agentic work stays in Codex; the explain/compare/draft overflow runs cardlessly in Telegram.