OpenAI Codex: Pay-As-You-Go, Plugins and the War with Claude Code

On April 2, 2026, OpenAI made one of its most aggressive launches of the year: Codex now has a pay-as-you-go model, a curated plugin directory, thread search, visual themes and -- most importantly -- two new code-optimized templates. And all of this comes at a time when Anthropic's Claude Code is dominating the development agent market.

This article analyzes each new feature, compares directly with what Claude Code offers today and helps you decide which tool to use for each scenario. No hype, no bias -- just facts, features and practical analysis.

1. What OpenAI Just Announced

The April 2026 Codex update package includes six significant changes that alter the way developers interact with the platform:

• Pay-as-you-go:You no longer need a Plus or Pro subscription to use Codex. Pay only for what you consume
• Curated plugin directory:an official marketplace where developers publish and install extensions for Codex
• Thread search:finally it is possible to search previous conversations by keywords, date and project
• Visual themes:interface costmization with light, dark and costm themes
• GPT-5.4:new flagship model with superior code reasoning performance
• GPT-5.3-Codex:model specifically optimized for coding tasks, faster than GPT-5.4 in code completions
• Windows sandbox:Native Windows support with proxy-only networking for security

Let's analyze each of these changes in depth.

2. Pay-as-you-go: how the new billing model works

Until now, to use Codex you needed a ChatGPT Plus ($20/month) or Pro ($200/month) plan. With pay-as-you-go, OpenAI eliminates this barrier: anyone with a free account can use Codex and only pay for the tokens consumed.

Pricing Structure

The charging model follows the same logic as the OpenAI API, but applied to the Codex environment:

Model	Input (for 1M tokens)	Output (per 1M tokens)	Optimal use
GPT-5.4	US$3.00	US$15.00	Complex tasks, refactoring, architecture
GPT-5.3-Codex	US$1.50	US$7.50	Fast completions, function generation, tests
GPT-4.1-mini	US$0.40	US$1.60	Simple tasks, formatting, short scripts

Who benefits

Pay-as-you-go is ideal for three profiles:

• Occasional users:Anyone who uses AI coding once or twice a week does not need to maintain a monthly subscription. Pay cents when you use
• Companies with large teams:Instead of purchasing 50 Pro licenses, the company can enable pay-as-you-go and only pay for each developer's actual consumption
• Who wants to test:the barrier to entry has dropped to zero. Create an account, deposit $5 and use Codex to decide if it's worth it before subscribing to any plan

What changes in practice

For most developers already using ChatGPT Plus, pay-as-you-go will probably not be the main mode. The subscription continues to be more economical for intense daily use. But the existence of this option removes an important objection: "I don't want to pay $20/month to test." Now you test for cents.

3. Plugin Directory: the ecosystem that OpenAI wants to build

This is probably the most strategic change in the package. OpenAI has launched a curated directory of plugins for Codex -- a marketplace where developers and companies publish extensions that add functionality to the agent.

How it works

Codex plugins are packages that integrate with the agent's execution environment. Each plugin adds specific capabilities:

• Infrastructure Plugins:deploy to AWS, GCP, Azure, Vercel, Netlify directly from Codex
• Database Plugins:connection and queries to PostgreSQL, MongoDB, Redis without leaving the agent
• Testing plugins:execution of test suites with integrated Jest, Pytest, Go test
• Design Plugins:conversion from Figma to code, generation of components from screenshots
• CI/CD Plugins:integration with GitHub Actions, GitLab CI, CircleCI

The curation model

Unlike the 2023 ChatGPT plugins (which failed due to lack of curation), OpenAI now manually reviews each plugin before publishing to the directory. There is a certification process that verifies security, performance and quality of documentation. This solves the "digital garbage" problem that killed the first attempt.

The directory launches with around 120 certified plugins, covering the most popular stacks: React, Next.js, Python/Django, Go, Rust, TypeScript and integration with the main cloud services.

Plugin monetization

Developers can publish free or paid plugins. OpenAI takes 30% of the revenue from paid plugins (the same model as the App Store). Prices range from US$5/month to US$50/month per plugin, depending on complexity.

4. GPT-5.4 and GPT-5.3-Codex: the engines behind

Models are the heart of any AI agent, and OpenAI has brought two new ones specifically to Codex.

GPT-5.4: the flagship model

GPT-5.4 is the direct evolution of GPT-5 (launched in late 2025). The most relevant improvements for coding:

• Longer chain reasoning:the model can maintain chains of thought of up to 128K tokens before starting to lose coherence -- compared to 64K for GPT-5
• Better understanding of large codebases:significantly better performance when dealing with repositories of more than 100 files
• Reduction of hallucinations in code:OpenAI reports a 40% reduction in "made-up code" (functions that don't exist, imaginary APIs) compared to GPT-5
• 256K tokens context window:twice as much as GPT-5, but still far from Claude Code's 1M

GPT-5.3-Codex: speed over depth

GPT-5.3-Codex is a smaller model, optimized for speed in code tasks. It is not ideal for systems architecture or complex refactoring, but it is excellent for:

• Function completions (2-3x faster than GPT-5.4)
• Generation of unit tests
• Quickly fix trivial bugs
• Code formatting and linting
• Converting between languages ​​for short snippets

In practice, most developers will use GPT-5.3-Codex for 80% of tasks and switch to GPT-5.4 when complexity demands. This dual model is a real competitive advantage.

5. Windows sandbox with proxy-only networking

Until now, Codex only ran natively on macOS and Linux. Windows support existed via WSL (Windows Subsystem for Linux), a functional but imperfect solution. OpenAI now brings native support to Windows with a different security approach: sandboxing with proxy-only networking.

What does this mean in practice

When Codex runs on Windows, it operates within a sandbox -- an isolated environment that does not have direct access to the user's operating system. All communication with the internet goes through a proxy controlled by OpenAI. This means:

• Security:the agent cannot execute arbitrary commands on Windows. If Codex tries to do something malicious (or makes a dangerous mistake), the sandbox prevents
• Controlled Networking:Codex only accesses the internet via proxy. You cannot make direct requests to external servers without going through the filter
• Predictable performance:the sandbox has allocated resources (CPU, RAM, storage), preventing the Codex from consuming too many machine resources

The difference in approach

Claude Code adopts a different philosophy: it runs directly on your system, without sandboxing, but asks for explicit permission before each action that changes something (creating files, executing commands, accessing the network). Safety comes from the consent model, not isolation.

Each approach has pros and cons. The Codex sandbox is more secure by default, but it limits the agent's flexibility. Claude Code's permissions model is more flexible, but it depends on the user paying attention to the authorizations. For corporate environments, sandboxing may be preferable. For individual developers who want full control, the Claude Code model is more productive.

6. Search for threads and visual themes

Two minor features that make a difference in everyday life.

Thread search

Until now, finding an old conversation in Codex was an exercise in frustration: infinite scroll, no filters, no search. The new functionality allows:

• Search for keywords in conversation content
• Filter by date (today, this week, this month, costm period)
• Filter by project (if you use workspaces)
• Search within the generated code, not just the conversation text

It seems basic, but anyone who accumulates dozens of threads a week knows how much this functionality was missing. Claude Code solves this in another way: the history is local to your terminal, and you can use system tools (grep, search) to find previous conversations.

Visual themes

Codex now supports costmizable visual themes: dark mode (default), light mode, and community-created themes. It's a cosmetic change, but professionals who spend 8+ hours a day on a tool value the ability to adjust the interface to their liking.

Claude Code, running in the terminal, inherits the theme of your favorite terminal emulator. If you use iTerm2 with the Dracula theme, Claude Code appears with these colors. Less visually polished, but infinitely more costmizable for those who already have a terminal setup.

7. Claude Code in 2026: Agent Teams, hooks and sub-agents

To contextualize the competition, it's worth looking at what Claude Code offers today that Codex doesn't yet have.

Agent Teams

Claude Code allows you to create "teams" of specialized agents who work in tollel on the same project. You can have one agent focused on frontend, another on backend, another on testing -- and they communicate with each other to deliver a cohesive result. This does not exist in the Codex.

In practice, an Agent Team works like this:

Hooks

Hooks are automations that fire before or after Claude Code's actions. Examples:

• Pre-commit hook:before each commit, Claude Code runs linting automatically
• Post-file-create hook:After creating a TypeScript file, it automatically generates the corresponding test file
• Pre-push hook:before pushing, run the entire test suite and only allow push if they all pass

Hooks transform Claude Code from a reactive agent (does what you ask) into a proactive agent (does things automatically based on rules you define). The Codex has nothing equivalent.

Sub-agents

Sub-agents are smaller agents that Claude Code can create and delegate tasks to. Unlike Agent Teams (which run in tollel), sub-agents are sequential: the main agent identifies that it needs a specialized task, creates a sub-agent with specific instructions, collects the result and continues.

For example: you ask Claude Code to create a landing page. It creates a sub-agent to generate the copy, another to write the CSS, and assembles everything at the end. Each sub-agent has an optimized context for its task, which improves the quality of the result.

1M Token Context

Claude Code operates with a context window of 1 million tokens -- 4x more than GPT-5.4's 256K. In practice, this means that Claude Code can "read" entire codebases of medium projects without having to discard information. In large projects, the difference between 256K and 1M tokens is the difference between the agent understanding the project partially or fully.

8. Direct comparison: Codex vs Claude Code

Time to put the two tools side by side. This table compares features, not opinions:

Functionality	OpenAI Codex (April 2026)	Claude Code (April 2026)
Main model	GPT-5.4 (256K context)	Claude Sonnet / Opus (1M context)
quick model	GPT-5.3-Codex	Claude Haiku
Billing model	Subscription + pay-as-you-go	Subscription or API credits
Extensions	Plugins (curated directory)	Skills (Markdown files)
Parallel agents	Nao	Agent Teams
Automations	Nao	Hooks (pre/post actions)
Sub-agents	Nao	Sub-agents
Interface	Web + desktop app	Terminal (CLI)
Security	Sandbox (Windows), permissions (Mac/Linux)	Explicit permissions model
Native Windows	Yes (sandbox)	Via WSL
Voicemod	Sim	Sim
History search	Yes (new functionality)	Via terminal (local grep)
Visual costmization	App themes	Terminal Themes
Access to local files	Sim	Sim
Command execution	Yes (sandbox on Windows)	Yes (directly in the system)

Where Codex Wins

• Accessibility:pay-as-you-go removes barriers. Graphical interface and more user-friendly for beginners
• Official plugins:curated directory offers ready-made and certified integrations
• Native Windows:Dedicated sandbox, no WSL required
• Thread search:finding old conversations is much easier

Where Claude Code wins

• Context:1M tokens vs 256K. For medium and large projects, there is no comparison
• Agent Teams:tollel agents working on the same project and a game-changer
• Hooks:automations that make the agent proactive, not just reactive
• Sub-agents:intelligent delegation of specialized tasks
• Skills:more flexible than plugins, easy to create and share
• Terminal performance:For devs who live in the terminal, nothing beats the speed of the CLI

9. Plugins vs skills ecosystem: which model wins?

This is the most interesting comparison in the long term, because it defines how each platform will evolve.

Codex Plugins

The Codex plugin model follows the logic of a traditional marketplace:

• Developers create plugins following a spec defined by OpenAI
• Submit for review (1-3 week process)
• If approved, they are available in the directory
• They can be free or paid (OpenAI takes 30%)
• Installation via graphical interface, with one click

Advantages:quality guaranteed by curation, simple installation, deep integration with the Codex runtime.

Disadvantages:slow process to publish, 30% fee, total dependence on OpenAI approval, proprietary format that only works on Codex.

Claude Code's Skills

Skills are Markdown files that you place in a folder and Claude Code reads them automatically. There is no marketplace, there is no approval, there are no fees:

• Anyone creates a skill: and a file.mdwith instructions
• Sharing via GitHub, direct download or packages like minhakills.io
• Installation: copy the file to the folder~/.claude/commands/
• No official curation (quality depends on the creator)
• Open format: the content of the skill is pure text, readable and editable by humans

Advantages:democratic, quick to create and share, no fees, open format, works offline.

Disadvantages:without official quality assurance, requires more knowledge to install (copy files), less native integration with external services.

In practice

For services and integrations (AWS, databases, CI/CD), Codex plugins have an advantage: the integration is deeper and more automated. For specialized knowledge (such as being an expert indigital marketing, SEO, tracking or a specific stack), Claude Code's skills are superior: you can create a skill that transforms Claude Code into any expert in minutes.

The skills model is more flexible. The plugin model is more polished. In the long term, the trend is for both to converge: plugins will become more open, skills will gain integration capabilities.

10. Which one to choose for what type of work

The honest answer: it depends on what you do. Here is a recommendation based on real scenarios:

Choose Codex if you:

• Works primarily on Windows and does not want to configure WSL
• Prefer graphical interface over terminal
• Uses AI coding sporadically and wants to pay per use
• Need deep integrations with cloud services via plugins
• Work in a team that is already in the OpenAI ecosystem
• Want integrated history search without having to use grep in the terminal

Choose Claude Code if you:

• Works on complex projects that require large context (1M tokens)
• Need tollel agents (Agent Teams) for maximum productivity
• Want automation via hooks for proactive workflows
• Lives in the terminal and values ​​CLI speed
• Need specialized skills for areas beyond code (marketing, design, management)
• Works with Mac or Linux natively
• Want full control over what the agent does on your system

Use both if you:

Many professionals are using both tools for different tasks. Codex for quick and specific tasks (taking advantage of pay-as-you-go), Claude Code for long and complex projects (taking advantage of the context of 1M and Agent Teams). There is nothing wrong with keeping both in the arsenal.

11. What to expect in the coming months

The war between Codex and Claude Code is heating up. Here's what we can expect based on current movements:

OpenAI

• Plugin directory expansion:from 120 to 500+ by the end of 2026. The more useful plugins, the greater the lock-in
• Codex Teams:It is very likely that OpenAI will launch its version of tollel agents to compete with Claude Code's Agent Teams
• Context boost:256K won't last. Pressure from Claude Code's 1M will force OpenAI to scale up
• Integration with IDE:native plugins for VS Code, JetBrains and other editors

Anthropic (Claude Code)

• Native Windows:Support via WSL will not last forever. Native support for Windows is a matter of time
• Skills marketplace:an official skills directory (curated or not) seems inevitable
• More integration with services:skills with the ability to connect to external APIs, approaching the plugin model
• Optional graphical interface:for those who prefer not to use a terminal, a web or desktop UI

The macro scenario

We are living in an arms race in AI coding. Codex has the user base of ChatGPT (hundreds of millions). Claude Code has the technical quality and the most advanced agent model. In the long term, the advantage goes to whoever builds the best extension ecosystem -- whether plugins or skills.

For the end user, the competition is excellent: both tools are improving every month, prices are falling, and capabilities are expanding. The best time to start using AI coding is now.