- 16
- May
2026 is the year AI coding agents became standard tooling for development teams worldwide — and from OpenAI's side, the hottest player is Codex, now back in two flavors: a CLI that runs in your terminal, and a Cloud agent that handles tasks asynchronously. This EP 1/4 lays the groundwork — what Codex is, how many forms it has, which models power it, and why enterprises are adopting it.
Quick Summary — What is Codex?
Codex is OpenAI's AI coding agent that returned in May 2025 after being deprecated in 2023. It now comes in two forms: Codex CLI (local, Rust-built, open source) and Codex Cloud Agent (delegate tasks to a sandboxed cloud), powered by GPT-5.5 and GPT-5.3-Codex. Included in ChatGPT Plus, Pro, Business, Edu, and Enterprise subscriptions.
What is Codex?
Codex is OpenAI's AI agent built specifically for software engineering. Unlike a typical autocomplete that completes lines, Codex acts as an "agent" — it reads existing code, analyzes context, edits multiple files, runs shell commands, executes tests, and produces pull requests.
What separates Codex from older AI coding tools is its autonomous capability. You can hand it a feature-level instruction like "add OAuth login," and it will modify multiple files, run tests, and return a complete PR for human review.
Timeline — From Codex v1 to Today
The name "Codex" isn't new, but the 2026 Codex is a completely different product from the original.
| Year | Event |
|---|---|
| 2021 | Original Codex launched — powered the first GitHub Copilot (autocomplete only) |
| Mar 2023 | OpenAI deprecated the original Codex API, folded it into GPT-3.5/GPT-4 — Codex disappeared from the market for over 2 years |
| May 2025 | OpenAI relaunched Codex as an AI agent — Codex CLI + Cloud Codex |
| Feb 2026 | Release of GPT-5.3-Codex, a coding-tuned model ($1.75 / $14 per 1M tokens) |
| May 2026 | Launch of GPT-5.5 — new frontier model for complex coding, computer use, and knowledge work — faster than GPT-5.4 and uses fewer tokens |
Codex CLI vs Cloud Codex — Two Forms You Must Know
Codex today ships in two modes. They work very differently — pick the wrong one and you get the wrong result.
| Attribute | Codex CLI | Cloud Codex |
|---|---|---|
| Runs where | Developer's machine (local) | OpenAI's cloud sandbox |
| Language | Rust (fast, lightweight) | Used via web UI or GitHub integration |
| License | Open source (Apache 2.0 on GitHub) | Proprietary — sign in with ChatGPT account |
| Best for | Interactive debugging, exploration, quick reviews | Batch tasks, long-running jobs, full feature delegation |
| Code access | Direct file access in working directory | Clones repo from GitHub into sandbox |
| Command execution | Runs shell on dev machine (with permission prompts) | Runs in isolated container — safer but harder to set up |
| MCP support | Yes — third-party tools via Model Context Protocol | Yes — but tools must be internet-accessible |
Models Powering Codex
Codex isn't a single model — users can switch models and adjust reasoning levels based on task complexity.
| Model | Best for | Strength |
|---|---|---|
| GPT-5.5 | Complex coding, refactor, debug, validation, knowledge work | Latest frontier model — fewer tokens than GPT-5.4, faster, higher quality |
| GPT-5.4 | General tasks (previous generation) | Stable, 272K-token context |
| GPT-5.4 mini | Light coding work, autocomplete | 2x faster than GPT-5.4, uses only 30% of quota |
| GPT-5.3-Codex | Code-specific tasks | Fine-tuned for coding — $1.75 in / $14 out per 1M tokens |
Model selection tip
For implementation or refactor work requiring high quality, use GPT-5.5. For autocomplete or light tasks, use GPT-5.4 mini to save quota. For code-specific tasks where per-token cost matters via API, use GPT-5.3-Codex.
Codex Highlights
- Switch model and reasoning mid-session — GPT-5.5 for the hard part, GPT-5.4 mini for the easy parts
- Code review by a second Codex agent — let another agent review the code before commit
- Subagents for parallel work — split a big task into pieces and have multiple agents tackle them concurrently
- Built-in web search — fetch current information from the internet mid-session
- MCP (Model Context Protocol) — plug in third-party tools like databases, observability, issue trackers
- Open source CLI — Codex CLI is open source under Apache 2.0 on GitHub; fork and customize at will
Why Codex Matters for Enterprises
Codex adoption growing through 2026 isn't a fad — there are solid business reasons:
- Bundled with subscriptions enterprises already buy — no separate license. ChatGPT Plus ($20/mo), Pro ($200/mo), Business, Edu, Enterprise all include Codex
- Cuts time-to-PR — work that used to take a day (writing tests for legacy code, refactoring a module) is done in an hour
- Parallel agents — a single team can have 3–4 PRs that Codex is working on simultaneously, like having multiple junior developers
- Less context switching — developers don't switch screens. Codex lives in the same terminal where the code is written
- Uses agentic AI architecture — Codex isn't just chat-and-respond. It runs a loop: plan → execute → verify → fix
Limitations to Know Before Adopting
Before deciding to use Codex in your organization, understand the limits:
- 272K-token context — smaller than Claude Code's 1M. Very large codebases may need to be chunked
- Sends code to OpenAI's cloud — enterprises with data residency or compliance constraints must evaluate first (we cover this in EP 4)
- All output must be verified — even with second-agent review, AI can still hallucinate. Never ship output without verification
- Weak on proprietary frameworks — if your organization uses an internal framework not on the public internet, the model doesn't know it. You'll need to provide documentation
- Some plans have usage caps — ChatGPT Plus has a monthly Codex cap. Power users may need Pro or credits
Who Should Use Codex?
Codex isn't an "AI that replaces humans" — it's a force multiplier for development teams. Best fit for:
- Small to mid-size engineering teams — boost PR velocity without hiring more
- Teams writing retroactive tests — cover legacy code that never had unit tests
- Developers running parallel work — delegate to Cloud Codex while doing other CLI work
- DevOps teams — automate scripts, refactor infrastructure-as-code, generate migration scripts
- Organizations using AI Cowork patterns — where dev-with-AI is the norm, not the exception
Summary — Codex EP 1/4
| Topic | Summary |
|---|---|
| What is Codex | OpenAI's AI coding agent, relaunched in 2025 — works as an agent, not just autocomplete |
| Forms | Two: Codex CLI (local, Rust, open source) and Cloud Codex (sandbox, async, GitHub integration) |
| Models | GPT-5.5 (latest frontier), GPT-5.4, GPT-5.4 mini, GPT-5.3-Codex (fine-tuned) |
| Pricing | Bundled with ChatGPT Plus, Pro, Business, Edu, Enterprise — API: GPT-5.3-Codex $1.75/$14 per 1M tokens |
| Best fit | Dev teams seeking velocity, legacy refactor, parallel task delegation, DevOps automation |
Codex doesn't replace developers — it turns one developer into a team of three or four. Organizations that adopt early gain a feature-shipping advantage.
- Saeree ERP Editorial
Continue Reading — EP 2, EP 3, EP 4
- EP 2: Codex vs Claude Code — 2026 AI Coding Tool Comparison
- EP 3: 8 Real-World Codex Use Cases + Limitations
- EP 4: Codex for Enterprises — License, Security, Governance
References
- OpenAI — Codex: AI Coding Partner
- OpenAI Developers — Codex CLI
- OpenAI Developers — Codex Pricing
- GitHub — openai/codex (open source CLI)
- OpenAI — Introducing GPT-5.5
If your organization is evaluating AI coding tools for the development team, the Saeree ERP team has experience deploying agentic AI in ERP systems for Thai enterprises. Schedule a consultation or contact our advisory team any time.
