AI・テクノロジー
Codex's Major Evolution | The Full Picture of the AI Developer That Automatically Controls Your PC
機械翻訳 / Machine-translated
A
AI Friends
2026年4月26日・22分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
From the era of "AI writes code" to the era of "AI operates your PC and gets work done on its own" —
On April 16, 2026, OpenAI delivered a major update to Codex, turning conventional wisdom about development on its head.
On your Mac screen, AI independently opens apps, clicks, and types — and integrates with more than 90 tools.
The popular tool used by 3 million developers per week is finally evolving to the level of a "dependable colleague."
Here's a thorough breakdown of what changed, in plain language anyone can understand.
Codex is a "developer-exclusive AI agent" that OpenAI is building alongside ChatGPT.
If ChatGPT is a "jack-of-all-trades tutor," then Codex is more like a "brilliant intern who specializes in programming."
It handles file operations, test execution, and Git operations all on its own.
In short, Codex has already reached a level of penetration where nearly 1 in 3 developers worldwide has used it. This update is disruptive precisely because it stacks a full suite of agent capabilities on top of that already widespread foundation.
Let's go through the headline features added in this update one by one.
The biggest highlight is "Computer Use," which lets Codex automatically operate your Mac in the background.
A dedicated AI cursor appears on your screen and performs clicks and keyboard inputs on its own.
And because multiple agents can run simultaneously, you can divide work three ways — "you handle email replies, one AI fixes code, another AI runs tests."
It's launching on macOS first, with the EU and UK expected to follow later. Windows hasn't been mentioned, but a gradual rollout is anticipated.
A dedicated browser is now built into the Codex app, letting you directly annotate web pages on display and give instructions like "change the color of this button" or "don't break the layout in this area." For front-end and game development, the ability to move back and forth between design and implementation is set to become dramatically faster.
Inside Codex, you can now use OpenAI's latest image model, "gpt-image-1.5," to generate and edit UI mockups and product concept images. Annotate a screenshot by hand with "add a blue button here," and the AI will redraw the image and update the corresponding code — all in one flow.
Codex now includes a "Memory Preview" that remembers your preferences, edit history, and past interactions.
No more re-explaining your coding habits every time — things like "always leave ESLint enabled" or "indent with 2 spaces."
Rolling out gradually to Enterprise, Edu, EU, and UK users.
Over 90 plugins have been added at once, making it possible to connect with virtually every tool developers use on a daily basis. Key highlights include:
codex marketplace add from GitHub, Git URLs, or local directoriesOn the technical side, MCP (Model Context Protocol) server support has been enhanced, with the ability to handle parallel calls and sandbox-state metadata.
One of the most-watched aspects of this update is the direct challenge it poses to Anthropic's Claude Code.
In short, the division of labor is becoming clear: "Claude Code for large-scale refactors reading hundreds of files," and "Codex for rapid test-and-debug cycles in the terminal." OpenAI itself has stated that "SWE-bench Verified is increasingly contaminated by training data, and SWE-bench Pro should become the new standard" — a sign that the benchmark war is heating up.
Starting April 2, 2026, Codex billing shifted from "per message" to "per API token." When usage approaches the limit, it's also possible to fall back to GPT-5.4-mini to keep going.
Individual developers and startups should start with Plus at $20/month and move up to Pro ($100/month) when they hit limitations.
Mid-sized teams and above who want to delegate large projects will find the gold-standard spring 2026 setup is using both Claude Max and Codex Pro together.
If you can only try one, starting with Plus to experience the new Computer Use feature is the smartest choice for cost-effectiveness.
Imagine a freelance web designer working in Setagaya, Tokyo.
Late at night, a client sends a request: "Change the CTA button on the landing page to red, set the margin to 30px, and make it two columns on mobile."
In the old days: wake up, make the changes manually, Git commit, deploy to Vercel — one hour of work.
Now: send instructions to Codex via Slack, and Computer Use opens a browser in the background, makes the edits, reviews, and deploys automatically.
The designer sleeps through it and just checks in the morning.
Take the case of a developer maintaining internal systems at a 50-person manufacturing company.
Codex reads 20 bug tickets stacked in Jira one by one via the Atlassian Rovo plugin, creates fix code, opens a PR, and runs tests on CircleCI — all autonomously.
The developer's job becomes just reviewing PRs and deciding whether to merge.
Daily output triples.
A startup product manager uses Codex + gpt-image-1.5 to create 3 mockup variations for a new feature before even involving a designer. Speaking naturally — "a bit more whitespace" or "give me a dark mode version too" — Codex updates the images and writes the corresponding React code, all in one continuous flow.
Yes.
As long as you have macOS, you can use it immediately from Japan.
Just log in to your ChatGPT account and download the Codex desktop app.
The interface supports Japanese, and instructions and responses in Japanese work without issue.
Note, however, that the Memory feature is rolling out to Enterprise, Edu, EU, and UK users first — users in Japan will need to wait a little longer.
Once Computer Use becomes widespread, the reality of "one engineer effectively doing the work of 3–5 people" will arrive.
Japan's contract development industry runs largely on the traditional model of headcount × hourly rate, and that model could unravel fast.
The shift to outcomes-based pricing and updating contracts to assume AI involvement should be priorities this fiscal year.
Domestic players like Preferred Networks and Sakana AI are also developing AI tools for developers, but breaking into the three-way dominance of Codex, Claude Code, and Cursor will not be easy. Differentiation through deep understanding of in-house codebases with Japanese comments, or specialization for regulated industries like finance and healthcare where compliance is paramount, will likely be the key.
A. It's designed to be fundamentally safe.
OpenAI runs agents in a sandbox environment, limiting them to apps and directories the user has explicitly permitted.
Sensitive operations like deletions or billing actions trigger confirmation prompts.
That said, it's best to avoid vague instructions like "do whatever you need in my Documents folder" — the recommended approach is to grant access on a per-project-folder basis.
A. As of this release, macOS only.
No timeline has been announced for Windows or Linux, but since the Codex CLI already runs on Linux, the other new features (plugins, memory, image generation) are expected to roll out across platforms. The Windows desktop app launched on March 4, 2026, so Computer Use is expected to follow within a few months.
A. For now, "use both" is the right answer.
Claude Code has the advantage on large-scale challenges like SWE-bench, but Codex is stronger for Terminal-Bench-style work and front-end tasks involving images and PC control.
If you can spend $200/month, subscribing to both and switching based on the task is the 2026 best practice.
A. The basics: get comfortable with Plus ($20/month), then upgrade to Pro ($100/month).
Switching to GPT-5.4-mini to stretch your quota is also an effective technique.
Through May 31, 2026, a campaign gives Pro users 10× the Codex quota of Plus — so subscribing to Pro before the end of May gets you maximum value.
A. The plugins themselves are free to use within your Codex plan, but the connected services (Jira, CircleCI, etc.) require separate subscriptions.
Using Atlassian Rovo requires a Jira license; using GitLab Issues requires a paid GitLab plan.
The smartest way in is to start with plugins that leverage services you're already paying for.
This major Codex update has a strong chance of becoming the watershed moment shifting AI from "writes code" to "does the work."
In a world where development productivity leaps 3× overnight, the human role will converge on two things: "deciding what to build" and "ensuring quality."
Right now, in spring 2026, is the perfect time to roll up your sleeves and try it.
In 30 minutes this weekend, you could reach out and grab a future where your work transforms dramatically.
This article is a cross-post from AI Friends.