Four million people have come in—what exactly are they using Codex for?

Article author and source: 0x9999in1, ME News

TL;DR

• Codex is no longer just a coding assistant. In May 2026, its weekly active users surpassed 4 million, with 50% of use cases no longer involving coding. It is evolving into a work platform capable of handling long-term tasks, retaining memory, executing periodic actions, and collaborating across tools.
• “Pinned thread + context compression” is a core paradigm shift. Work is no longer “one question, one answer,” but rather a persistent conversation space that can run for months, accumulating history, preferences, and decision records.
• Heartbeats (Automated Heartbeats) teaches Agents to "not wait." It can check Slack and Gmail every 30 minutes, monitor a feedback thread every 15 minutes, and even queue for customer service and claim refunds on your behalf.
• The key to a memory system isn't "remembering"—it's "writing it down." Serialize valuable information from threads into files using an Obsidian vault + Git diff to create long-term, reusable knowledge assets.
• Remote control and mobile review truly free your work from the desk. Starting May 14, both iOS and Android users can control Codex sessions, review outputs, approve commands, and adjust direction via the ChatGPT mobile app.
• Goals represent the transition from "prompting" to "assigning." With clear validation criteria, the Agent can autonomously run in a continuous loop for hours, even up to 14 hours.
• The sidebar is the key interface that makes Codex the place where work happens. It’s not just a preview of the output—it’s a unified space for inspecting, annotating, interacting with, and reviewing changes.

A number is placed here first

4 million.

This is the number of weekly active developers for Codex, as reported by OpenAI at the end of April 2026. Two weeks ago, it was 3 million. Going further back, between January and April 2026, the number of Codex users in ChatGPT Business and Enterprise increased sixfold.

How fast is it? In its first week, the Codex desktop app surpassed 1 million downloads. By the time of the Codex App’s launch in mid-May, OpenAI even temporarily granted access to all ChatGPT Free and Go users, and doubled the rate limits for Plus and Pro users.

This is not the growth curve of a niche tool within the programmer community. This is the growth curve of a product that is consuming work itself.

Here's the question: With four million people coming in, what are they actually using Codex for?

The answer may surprise many—50% of users no longer use it for coding.

Who is using it, and how?

Jason Liu is an engineer on the OpenAI Codex team. On May 17, he published a personal blog post titled "Codex-maxxing," documenting how he transformed Codex from a programming tool into a comprehensive knowledge work platform.

This article received 8,800 views and 225 saves within 24 hours of being posted on X, with 100% positive sentiment in the comments. It is not promotional content, but a genuine user’s detailed daily log.

His usage made me realize one thing: most people are still using Codex to “ask questions,” while those who truly leverage it are already using it to “run workflows.”

What's the difference?

The question is: I give you a prompt, you give me an answer, the conversation ends, and the task is interrupted.

The workflow is: I give you space to keep moving forward, and I can return at any time to check in, guide, or make corrections—tasks never need to start from scratch.

This is what Codex is truly selling in 2026—not better code generation, but workflow continuity.

Persistent threads: Not chat logs, but workstations

The first feature to change everything is called context compression.

Traditional AI conversations have a critical flaw: the context window is limited. The longer the conversation, the more information is lost, eventually turning the assistant into one with amnesia. Codex’s solution is context compression—allowing threads to extend infinitely while preserving core information.

Jason Liu's approach is to maintain a pinned thread for each important workflow.

He has a "Chief of Staff" thread dedicated to sorting inbox and Slack messages. There is an "Agents SDK" thread tracking progress on open-source projects. A separate "Open Source" thread manages community contributions. There is also a dedicated thread monitoring Twitter.

These are not short conversations. They are massive threads that have been running for months, accumulating extensive history, preferences, and past decisions.

Use the keyboard shortcuts Command-1 through Command-9 to jump directly.

You might be wondering: Isn't this just making the chat log longer?

No. The length isn’t the point. The point is that these threads begin to operate like “workstations”—each with its own domain, its own memory, its own rhythm. You’re no longer “chatting with an AI”; you’re “returning to an ongoing task.”

Of course, there is a cost. Once a long-running thread is evicted from the cache, the cost to reload it is higher. But for truly critical workflows, the value of continuity far outweighs the cost.

Voice input + Steering: Let the Agent grasp your true intentions

Voice input in Codex is not a feature for "convenient typing." Its true value lies in giving the Agent your unedited thoughts as they are.

Jason Liu gave an example. If he were planning his work and said, “I remember someone named Ben mentioned this in Slack, but I can’t recall the details—can you look it up?”—typing it out would sound vague and annoying. But saying it aloud feels very natural. And an Agent can handle such vague instructions perfectly.

More interesting is the Steering feature. It allows you to continuously add intent while the Agent is working. For example, while reviewing a website, you can say as you go:

• Reduce this a bit.
• This copy is incorrect.
• The spacing feels uncomfortable
• Open a PR when done.
• Once deployment is complete
• Send the preview link to Slack

You don’t need to wait for each step to finish before deciding the next one. You can queue up everything you plan to do next while the Agent is still executing, then walk away.

What has changed? The relationship between people and Agents. You are no longer a user who can only send one message at a time and wait for a reply. You’ve become a commander who can continuously input intentions, allowing the system to queue and execute them on its own.

Memory: Not remembering the conversation, but writing it to a file

Once a thread becomes long-lived, a new issue emerges: how is knowledge shared across threads?

Codex has a first-party memory feature (Settings > Personalization > Memories), but Jason Liu goes further: he uses an Obsidian vault as the Agent's "home":

vault/
├── TODO.md
├── people/
├── projects/
├── agent/
└── notes/

At the top level in AGENTS.md, specify: When you learn more about someone, advance a project, or close a to-do loop, update the corresponding page in the vault.

Then maintain this vault as a GitHub repository.

Why? Two reasons.

First, it can run in the cloud without being tied to a local environment.

Second—and this is critical—diff becomes the interface for reviewing memory.

When the Agent updates the vault, you can read the Git diff to see what information it considers important enough to remember. This review step is crucial. You don’t want a long-running thread to quietly accumulate a vague “feeling” in the conversation history. You want it to document what has actually changed: someone’s preferences, what a project is waiting for, a decision that has been made, or a cycle that has been closed.

The purpose of a memory system is not to "remember more," but to compress experience into a form that can persist independently of the thread. If the thread disappears or the compression is poor, the useful knowledge remains in the files.

Chronicle is another memory solution launched by OpenAI on April 20, building contextual memory by capturing screen content in the background. Sam Altman described the experience as "like telepathy." However, it involves real trade-offs: screenshots are processed through OpenAI servers and stored locally as unencrypted Markdown files, increasing the attack surface for prompt injection. It is currently a research preview feature that requires manual activation.

The direction is correct, but it’s not yet mature. The file system remains the most reliable memory infrastructure.

Heartbeats: Enable the Agent to work proactively

This may be Codex's most disruptive feature.

Heartbeat is a thread-local scheduled automation. You can tell a thread: “Check on this every few hours,” and it will run automatically at that interval.

Jason Liu's Chief of Staff thread runs every 30 minutes:

Check Slack and Gmail every 30 minutes for messages that require attention but haven't been replied to. Help me prioritize them. If someone asks me a question, research the answer as thoroughly as possible and draft a response, but do not send it.

When he returned to Slack, many replies were already in draft. The most labor-intensive context collection was complete. He just needed to decide which ones to send.

Another example is more vivid: he was working on an animation project, posted the video on Slack for colleagues to review, and had Codex check the thread every 15 minutes. If there were comments, it would re-render a new version and reply. Since the Slack MCP server doesn’t support file uploads, the Agent used the @computer capability to click the “Add file” button and upload the file.

Notice what’s happening here: a feedback loop spans across three tools—Slack (collecting feedback), Remotion (rendering animations), and @computer (uploading files). The agent maintained this loop on its own after the person left.

Here’s an even more extreme example: His package was stolen, and Amazon said he had to wait 25 minutes to reach a human agent. He created a thread tagged with @computer: checking every 5 minutes if a客服 had joined, and if so, pushing for a refund—then increasing the frequency to once per minute. He then went to take a shower. When he came out, the refund had already been processed.

This is not science fiction. This is a real-world use case taking place in May 2026.

Remote control: Work no longer tied to your desk

On May 14, OpenAI officially rolled out Codex to the ChatGPT mobile app, available on both iOS and Android. All ChatGPT plans are supported, including Free and Go.

TechCrunch's report quoted OpenAI's statement: "This is not just about remotely controlling individual tasks or assigning new ones. From your phone, you can work across all threads, review outputs, approve commands, switch models, or initiate new tasks."

Execution and credentials remain on the desktop machine. The mobile device is only responsible for review, approval, and guidance.

This forms a complete workflow together with Heartbeats and pinned threads:

1. Start a long-running task on your desktop.
2. Heartbeat makes it run automatically in rhythm.
3. You step away from your desktop and check the progress on your phone.
4. Intervene to guide at key decision points
5. Work continues to progress

Work doesn't pause just because you've changed locations. This is true asynchronous workflow.

Goals: From prompting to assigning

Goals were introduced in Codex CLI 0.128.0/goal command. It enables the Agent to enter an autonomous loop: plan → act → test → review, until your defined success criteria are met or the quota is exhausted.

A weak goal is: "Execute the plan in this Markdown file."

A strong goal is: “Migrate the Python Rich library to Rust and pass all unit tests of the original Python library.”

The latter has genuine verification criteria. The Agent can independently determine whether the task is complete, where it failed, and how to fix it.

Someone used /goal to have Codex run continuously for 14 hours on a device driver project. Someone else used it to build a complete game in under an hour.

What are the key differences?

If you give an agent only a vague plan, it’s at best a hardworking intern. If you give it a goal with clear validation criteria, it becomes a self-correcting executor.

No verified ambition, just desire. This applies to agents as well as to people.

Sidebar: Where the work happens

Many people think of the side panel as a "preview window," but this greatly underestimates its capabilities.

The sidebar does three things:

Review outputs. Markdown is commentable, spreadsheets render formulas and support cell editing, CSV displays as a table, PDF renders directly (a blessing for LaTeX users), and slides can be reviewed in-app. The key isn’t just that Codex can generate these—it’s that you can review and annotate them without breaking your flow.

Interact with the web interface. The Agent can control the in-app browser using $browser with JavaScript and can directly add annotations on the rendered content. The minimal and most persistent version is a single-file index.html with JS and CSS—no server required, open and use immediately. Storybook for component review, Remotion Studio for animations, Slidev for presentations, and Streamlit for data applications can all be accomplished within this space.

Review changes. Diff, PR comments, and code modifications—all can be completed in the same interface.

When the content on the sidebar and the object the Agent is acting on are the same thing, your collaboration efficiency transforms dramatically. You no longer need to switch back and forth between the chat window and the actual output—the output becomes part of the conversation itself.

Positioning and Pricing: What is it selling?

On April 2, 2026, OpenAI changed the pricing for Codex from per-message billing to API token usage billing.

The current structure is:

• Codex is included in all ChatGPT plans (Free, Go, Plus, Pro, Business, Enterprise)
• Enterprise users can add 'Codex-only seats' and pay based on usage, with no fixed seat fees.
• OpenAI estimates that the average developer spends $100–200 per month.

New users can receive up to $500 in credit when registering as a team.

This pricing strategy is clear: get users started first, then charge based on usage. Once your workflows are running on Codex—threads have accumulated memory, Heartbeats are running in the background, and context has been stored in vaults—the cost of migration becomes extremely high.

This is not the logic of SaaS. This is the logic of infrastructure.

What sets us apart from competitors?

By 2026, the AI programming tools market has split into three paradigms:

• Claude Code: Native terminal, largest context window (200K tokens standard, 1M beta), deep code reasoning
• Cursor: Native IDE, real-time inline editing, fastest for daily coding
• Codex: Cloud-based asynchronous, fire-and-forget mode, long-running tasks + multi-tool orchestration

They are not competing on the same level.

Cursor wins with its feel. As you type a line of code, it instantly completes it—so fast it’s addictive. Over 360K paying users have proven it.

Claude Code excels in depth. For a large task requiring refactoring across 50 files, its context understanding may be the best.

Codex wins through continuity.

It’s not the fastest tool for writing code right now. It’s the tool that keeps working even after you’ve walked away. Heartbeats, remote control, Goals, memory system, Connectors—these features together don’t create a better editor; they create a continuously running workflow operating system.

This is a completely different product positioning.

The real transformation

Let me put these pieces together.

The past AI usage pattern was like this:

User requests → Model provides result → Conversation ends → Task interrupted → Resume next time.

The model Codex is building is as follows:

User sets goal → Thread runs continuously → Memory is saved to file → Heartbeats execute periodically → User can intervene at any time to review → Work never resets to zero.

From "question and answer" to "running cycle." From "capability" to "continuity." From "chat window" to "operating system for work."

The significance of this shift is not that agents become smarter, but that the way work is organized has changed.

In the past, it was common for an idea to die after the prompt. You had an inspiration, opened ChatGPT, asked a question, received an answer, and then closed the window. Three days later, you had already forgotten about it.

Now, an idea can be placed into a thread, equipped with memory, set with a heartbeat, and attached to verification criteria, then continuously progressed. You don’t need to watch it constantly—just check in occasionally to review the diff and ensure the direction hasn’t drifted.

This is not about replacing human work. It’s about preventing more work from dying halfway.

For knowledge workers, this could be the key step for AI tools to truly integrate into daily workflows—not because they’ve become stronger, but because they’ve finally learned to never clock out.

Reference sources

1. OpenAI, "Scaling Codex to enterprises worldwide," openai.com, April 2026.
2. OpenAI, "Codex for (almost) everything," openai.com, May 2026.
3. OpenAI, "Introducing the Codex app," openai.com, May 2026.
4. Jason Liu (@jxnlco), "Codex-maxxing," jxnl.github.io/blog, May 10, 2026.
5. TechCrunch, "OpenAI says Codex is coming to your phone," techcrunch.com, May 14, 2026.
6. OpenAI, "Codex now offers pay-as-you-go pricing for teams," openai.com, May 2026.
7. OpenAI, "Top 10 uses for Codex at work," openai.com/academy, May 2026.
8. 9to5Mac, "Codex for Mac gains Chronicle for enhancing context using recent screen content," 9to5mac.com, April 20, 2026.