OpenAI Codex and Anthropic’s Claude Code are, in my view, two of the most important AI coding agents today. How do they differ in workflow, features, models, security, context, and cost?

Note (August 2026): Parts of this post are now outdated. In my hands-on comparison Claude Code vs. Codex vs. Cursor I revise, among other things, the assessment made here that Codex feels “more polished and more mature” overall. The model versions mentioned (GPT-5.4, Opus 4.6) are no longer current either.

Introduction

OpenAI Codex and Anthropic Claude Code are currently among the most important AI coding agents. Both offer extensive capabilities for a professional AI-driven development workflow. In this post, I compare the two tools and assess them based on my day-to-day experience using them in real projects.

OpenAI Codex

Overview

Codex is OpenAI’s coding agent. It is available as a command-line tool, as a plugin for several IDEs, and as a standalone desktop app. It supports both local and cloud-based workflows and multiple models, including GPT-5.4. I already introduced Codex some time ago in my post OpenAI Codex vs. GitHub Copilot. Since then, however, quite a lot has changed. So it makes sense to take another closer look at Codex and compare it to its direct competitor, Claude Code.

Features

Regardless of whether you use Codex via the CLI, an IDE plugin, or the Codex app, the tool mainly offers the following features:

• Security: Compared to its competitors, Codex follows a very conservative security model with restrictive execution modes and an operating-system-level sandbox. Depending on the selected configuration, filesystem and network access are restricted, and even approved commands do not necessarily run with full host access. In open modes such as full access, however, these safeguards only apply to a limited extent or not at all. In practice, Codex significantly reduces the risk of dangerous commands such as rm -rf / in its default configuration.
• Tools: Codex supports the installation of so-called “skills” and “plugins.” These are predefined actions or workflows the agent can execute. There are plugins for GitHub, Jira, and Figma, for example. That allows the agent to create pull requests, open Jira issues, or analyze and implement designs from Figma.
• Multi-agent support: Codex can run multiple agents in parallel. The main agent delegates subtasks to subagents. This improves efficiency and reduces waiting times. It also helps reduce “context pollution,” meaning the context gets less cluttered with irrelevant information that might hurt output quality. A typical example is analyzing the codebase before the actual coding starts. Instead of letting the main agent do that work, Codex can assign it to a subagent that extracts and summarizes the relevant information. That keeps the main agent’s context cleaner and more focused on the overall task. In practice, this works surprisingly well.
• Models: Codex can be used with different OpenAI models that vary in capability, latency, and cost. I use the newest and most capable model almost exclusively. As of mid-April 2026, that is GPT-5.4. In my experience, it produces the best results while still remaining cost-effective enough. The time lost through poor or mediocre results from cheaper models would be more expensive in practice than paying extra for the flagship model.
• Context window: Codex offers a context window of around 272,000 tokens as a hard limit. The total number of input and output tokens in a prompt must not exceed that limit. In practice, though, that is only part of the story, because models only produce consistently good results within a certain portion of the full context window, for example up to roughly 75 percent. If you want to dig deeper into that topic, the MRCR score is a useful concept to look at because it quantifies this effect at least partially. OpenAI does not publish concrete numbers for this “effective context size,” but it is reasonable to assume that it is somewhere around 200,000 to 220,000 tokens. Very roughly, that corresponds to about 12,000 to 15,000 lines of code, depending on the language and formatting. That number is reduced further by prompt guidelines, such as a bootstrap prompt or architecture documentation, and by the number of configured tools. Even so, Codex can usually keep around 50 to 100 source files in view for a typical coding task, which is sufficient for most real-world use cases.

Pricing

Codex is included in all OpenAI subscriptions starting with the Plus plan. The number of allowed requests increases with each subscription tier. If you hit a limit, OpenAI allows you to purchase additional “credits.” In my opinion, OpenAI is still very generous with Codex rate limits, at least as of April 2026 and probably until the end of May 2026. Even with the cheapest Plus plan, it is possible to use Codex in a professional AI-driven development workflow without constantly running into the limits. I assume, however, that this will change in the medium term. Based on the subscription changes that have recently become known, OpenAI is aligning its plans more strongly with heavy Codex usage, including an additional Pro tier between Plus and the previous Pro tier. In the end, it is probably not realistic to expect nearly unlimited access to an almost perfectly functioning AI agent for less than EUR 20 per month when that agent can easily handle the work of a junior developer.

Claude Code

Overview

Claude Code is Anthropic’s counterpart to OpenAI Codex and has been available since early 2025, initially only as a command-line tool. Since around the end of 2025, Claude Code has also been available, much like Codex, via IDE plugins and a separate desktop app. The current models, especially Claude 4.6 in its Opus and Sonnet variants, are very strong in the coding domain. Anthropic recognized the importance of AI-driven development early and has built a strong market position as a result.

Features

Claude Code offers the following key features:

• Security: Claude Code uses a configurable permission system for tools. With allowlists and denylists, you can define in a very granular way which actions may be executed automatically and which require manual confirmation. Unlike Codex, however, tools are not executed in a true operating-system-level sandbox by default. Anthropic does now offer OS-level sandboxing as well, but it has to be enabled explicitly and is not the default setting.
• Tools: Like Codex, Claude Code can connect to external services via the Model Context Protocol (MCP).
• Multi-agent support: Claude Code supports running subagents in parallel. The idea is the same as in Codex: more efficiency through parallelization and cleaner contexts through separation of concerns. Each subagent only receives the tools relevant to its task, which has a positive effect on both output quality and cost.
• Models: As of April 2026, Claude Code offers three model tiers: Opus 4.6 for complex implementation and architecture tasks, Sonnet 4.6 for everyday coding work, and Haiku 4.5 for fast requests, although I have not personally used Haiku yet. In my view, Opus 4.6 is still slightly more capable than GPT-5.4. It follows instructions very well and produces code at a level that only a small number of developers in human teams can consistently match.
• Context window: The usable context size depends on the selected subscription. In the cheapest subscription tier, the maximum available context is 200,000 tokens, which is slightly less than Codex offers. Starting with the Max plan, Claude Code provides a context window of 1,000,000 tokens for Opus 4.6, which is roughly 3.5 times the standard context window of Codex. As with Codex, the effectively usable context size is somewhat smaller in practice. My rough estimate is that Claude Code can then process around 40,000 to 50,000 lines of code, or about 150 to 350 files, efficiently and reliably. In complex projects with many modules, services, and interfaces, that larger context window is a real advantage. Of course, you also pay for that advantage with higher costs, as discussed below.

Pricing

Claude Code is included in all paid Anthropic plans. As already mentioned, the large context window is only available from the cheapest Max plan onward without additional usage billing. As of April 2026, that is about EUR 100 per month. The Pro plan, the cheapest plan for individual users, is roughly on par with OpenAI’s Plus plan in terms of price and context window. That does explicitly not apply to rate limits, though: while OpenAI currently offers very generous limits for Codex, Claude Code’s limits are much stricter. That noticeably reduces the practicality of using Claude Code’s Pro plan for professional AI-driven development. In my experience, the default limit is enough for at most one moderately complex feature per day, which is simply not sufficient. To Anthropic’s credit, it has recently added the option to increase limits through additional purchases without requiring you to switch straight to the Max plan. Even so, I think EUR 100 per month for the cheapest Max plan is still entirely reasonable given Claude Code’s capabilities.

Conclusion

I have been using both OpenAI Codex and Claude Code for several months in my daily AI-driven development workflow. Both tools perform extremely well in practice. When it comes to the quality and feature set of the IDE plugins and desktop apps, OpenAI currently has the edge in my view. Codex simply feels more polished and more mature overall. Anthropic’s Opus models, on the other hand, are still slightly ahead of GPT-5.4 when it comes to pure coding.

At this point, I see hardly any relevant task in day-to-day AI-driven development where Codex or Claude Code could not be used productively. That applies both to pure coding tasks and to conceptual questions, such as discussing architecture and design decisions.

In practice, I regularly switch between the two tools depending on the task at hand. I like to hand very complex implementation tasks to Claude Code and Opus 4.6, while I use Codex with GPT-5.4 for most everyday coding tasks, conceptual discussions, and documentation. For reviews, I often let both agents look at the same issue in parallel to get multiple perspectives. In my view, it is neither necessary nor particularly sensible to commit to a single tool for an AI-driven development workflow. In a professional environment, there is nothing wrong with using both tools in parallel and choosing whichever one is better suited to the task, even if that means taking on some additional cost.