Why is OpenAI actually chasing Claude Code?

Inside OpenAI’s Race to Catch Up to Claude Code

By Maxwell Zeff, Wired

Translated by Peggy, BlockBeats

Editor’s Note: Amid the rapid rise of AI programming agents, OpenAI—once leading the generative AI wave with ChatGPT—has unexpectedly become a “follower” in this critical race. In stark contrast, Anthropic, founded by former OpenAI members, has quickly gained popularity in the developer community and enterprise market with Claude Code, becoming one of the leading AI programming tools.

Through interviews with OpenAI executives, engineers, and multiple developers, this article reveals the true story behind the competition: from the early split of the OpenAI Codex project and shifting resources to ChatGPT and multimodal models, to internal team reorganization and accelerated release of AI coding products, OpenAI is experiencing a turning point—from strategic neglect to full-scale catch-up. In a sense, this is not a technical capability lag but a strategic pacing mismatch: ChatGPT’s explosion changed company priorities, Microsoft’s partnership limited product pathways, and Anthropic bet earlier on the AI coding track.

Deeper issues are also emerging: as AI agents begin to undertake more cognitive tasks, the software development process—and even white-collar work—may be redefined.

Below is the original article:

OpenAI CEO Sam Altman sits with his legs crossed on his office chair, gazing at the ceiling as if pondering an as-yet-formed answer. To some extent, this also relates to the environment.

OpenAI’s new headquarters in Mission Bay, San Francisco, is a modern building made of glass and light wood, almost like a “tech cathedral.” On the display shelves behind the reception, there are manuals introducing the “Eras of AI,” as if depicting a path to technological enlightenment. The staircase walls are covered with milestone posters of AI development, including one marking the moment when thousands of viewers witnessed a machine defeat top esports teams in Dota 2 via live stream. In the corridor, researchers in team-branded shirts with slogans like “Good research takes time” come and go. Of course, ideally, not too long.

We are seated in a large conference room. The questions I pose to Altman relate to the AI programming revolution sweeping the industry, and why OpenAI seems not to be leading this wave.

Today, millions of software engineers are starting to delegate some coding tasks to AI, forcing many in Silicon Valley to confront a stark reality: automation might impact their own jobs. Coding agents have thus become one of the few applications where companies are willing to pay high prices for AI. Logically, this moment could—and perhaps should—be OpenAI’s next “victory” headline. But now, the headline is not about OpenAI.

The company’s competitor is Anthropic, an AI firm founded by former OpenAI members. With its product Claude Code, Anthropic has experienced explosive growth. The company disclosed in February that this product accounts for nearly one-fifth of its business, with annualized revenue exceeding $2.5 billion. In comparison, according to an insider, by the end of January, OpenAI’s own coding product, OpenAI Codex, had an annualized revenue just over $1 billion.

The question is: why is OpenAI falling behind in this AI coding race?

“The advantage of being first is very significant,” Altman said after a moment of reflection. “We’ve already experienced this with ChatGPT.”

However, he believes now is the time for OpenAI to fully leverage AI coding. He thinks the company’s existing models are powerful enough to support highly complex coding agents. Of course, such capabilities are not accidental; the company has invested billions of dollars into training these models.

“This will be a huge business,” Altman said. “Not only because of the direct economic value, but also because of the productivity gains that coding can unleash.” He paused, then added, “I rarely use this word lightly, but I think this could be one of those markets worth tens of trillions of dollars.”

Furthermore, he believes that OpenAI Codex might be the “most likely path” toward Artificial General Intelligence (AGI). According to OpenAI’s definition, AGI is an AI system capable of surpassing human performance in most economically valuable tasks.

Sam Altman, CEO of OpenAI. Photo: Mark Jayson Quines.

Despite Altman’s calm and confident stance, the internal reality over the past few years has been much more complex. To gain a fuller picture, I interviewed over 30 insiders, including current OpenAI executives and employees who spoke on condition of anonymity, as well as former staff who shared insights into the company’s internal workings. Taken together, these accounts reveal an uncommon situation: OpenAI is racing to catch up.

Rewinding to 2021, Altman and other OpenAI executives invited Wired’s Steven Levy to their early office in San Francisco’s Mission District to witness a new technology demo. It was a project based on GPT-3 derivatives, trained on大量开源代码来自 GitHub.

During the demo, executives showcased how the tool, called OpenAI Codex, could accept natural language instructions and generate simple code snippets.

“It can actually perform operations in the computer world for you,” explained Greg Brockman, OpenAI’s President and co-founder. “You have a system that can truly execute commands.” Even then, OpenAI researchers generally believed Codex would be key to building a “super assistant.”

At that time, Altman and Brockman’s schedules were almost entirely filled with meetings with Microsoft—OpenAI’s largest investor. Microsoft planned to use Codex to power one of its first commercial AI products: GitHub Copilot, a code completion tool embedded directly into developers’ IDEs.

An early OpenAI employee recalled that at that stage, Codex “basically only did autocomplete.” But Microsoft executives still saw it as a significant signal of the AI era’s arrival.

In June 2022, when GitHub Copilot was officially released, it attracted hundreds of thousands of users within months.

Greg Brockman, President of OpenAI. Photo: Mark Jayson Quines.

The team responsible for Codex was later reassigned to other projects. An early employee recalled that the company’s view was: future models would inherently have coding abilities, so maintaining a dedicated Codex team long-term was unnecessary. Some engineers moved to work on DALL-E 2, others shifted to training GPT-4. At the time, this was seen as a key step toward closer proximity to AGI.

Then, in November 2022, ChatGPT launched and gained over 100 million users in two months. Almost all other projects within the company were paused. In the following years, OpenAI did not have a dedicated team for AI coding products. A former member involved in Codex said that after ChatGPT’s popularity, AI coding no longer fit into the company’s new “consumer product first” strategy. Meanwhile, industry consensus held that the field was “covered” by GitHub Copilot, which was essentially Microsoft’s turf. OpenAI was mainly providing underlying models.

Thus, in 2023 and 2024, OpenAI shifted more resources toward multimodal AI models and intelligent agents. These systems are designed to understand text, images, videos, and audio simultaneously, and operate like humans—controlling cursors and keyboards. This direction seemed more aligned with industry trends: Midjourney’s image generation models went viral on social media, and industry experts believed that large language models must be able to “see” and “hear” the world to reach higher levels of intelligence.

In contrast, Anthropic took a different path. While also developing chatbots and multimodal models, it appears to have recognized the potential of coding capabilities earlier. In a recent podcast, Brockman admitted that Anthropic had “focused heavily on coding ability” from an early stage. He noted that during training, Anthropic used complex programming problems from academic competitions, as well as大量来自真实代码仓库的“混乱”代码。

“This was a lesson we only realized later,” Brockman said.

In early 2024, Anthropic began training Claude 3.5 Sonnet using data from these real code repositories. When the model was released in June, many users were impressed by its coding skills.

This performance was especially validated by a startup called Cursor, founded by a group of young people in their twenties. They developed an AI coding tool that allows developers to describe requirements in natural language, and the AI directly modifies the code. After integrating with Anthropic’s new model, Cursor’s user base grew rapidly, according to an insider close to the company.

A few months later, Anthropic started internal testing of its own coding agent product, Claude Code.

As Cursor’s popularity soared, OpenAI attempted to acquire the startup. But multiple sources close to Cursor said the founding team rejected the deal before negotiations deepened. They believed the AI coding industry had huge potential and wanted to remain independent.

Andrey Mishchenko, head of OpenAI Codex research. Photo: Mark Jayson Quines.

At that time, OpenAI was training its first “reasoning model,” called OpenAI o1. These models could perform step-by-step reasoning before providing answers. OpenAI claimed at release that the model excelled at “accurately generating and debugging complex code.”

Mishchenko explained that a key reason for the significant progress in AI coding ability was that programming is a “verifiable task.” Code either runs or it doesn’t, providing very clear feedback signals. When errors occur, the system can quickly identify where the problem is. OpenAI leveraged this feedback loop to train o1 on more complex programming problems.

“If you don’t have the ability to explore freely in code repositories, implement modifications, and test your own results—these are part of ‘reasoning’—then today’s coding agents cannot reach their current level,” he said.

By December 2024, multiple small teams within OpenAI had begun focusing on AI coding agents. One team was led by Mishchenko and Thibault Sottiaux, who previously worked at Google DeepMind and now he is head of OpenAI’s Codex.

Initially, their interest in coding agents stemmed from internal R&D needs—aiming to automate repetitive engineering tasks—such as managing model training jobs and monitoring GPU clusters.

Another parallel effort was led by Alexander Embiricos, who previously managed OpenAI’s multimodal agent projects and now is product lead for Codex. Embiricos had developed a demo called Jam, which quickly gained internal attention.

Thibault Sottiaux, head of OpenAI Codex. Photo: Mark Jayson Quines.

Unlike controlling a computer via mouse and keyboard, Jam could directly access the command line. The 2021 Codex demo only showed AI generating code for humans to run manually; Embiricos’s version could execute the code itself. He recalls watching a live webpage recording Jam’s actions refresh on his laptop, almost stunned.

“For a while, I thought multimodal interaction might be the way to achieve our mission. Like humans sharing screens and working together all day,” Embiricos said. “Then it suddenly became very clear: maybe giving the model direct programmatic access to the computer is the real way.”

These scattered projects took months to integrate into a unified direction. By early 2025, when OpenAI finished training OpenAI o3—a model optimized further for coding tasks—the company finally had the technical foundation to build true AI coding products. Meanwhile, Anthropic’s Claude Code was ready for release.

Before Claude Code’s release (initially as a “limited research preview” in February 2025, fully launched in May), the dominant mode in AI coding was called “vibe coding.” Developers used AI-assisted tools to push projects forward, with humans controlling the direction and AI providing specific implementations. These tools had attracted hundreds of millions of dollars in investment.

But Anthropic’s new product changed this paradigm. Like the Jam demo, Claude Code could run directly via the command line, meaning it could access all developer files and applications. Coding was no longer just “AI-assisted”; developers could delegate entire tasks directly to AI agents.

Faced with this shift, OpenAI accelerated the release of competing products. Sottiaux recalled that in March 2025, he assembled a “sprint team” to integrate multiple internal groups and quickly launch an AI coding product.

Meanwhile, Altman also sought to “overtake on the curve” through acquisitions, offering $3 billion to buy AI startup Windsurf. Executives believed this deal would bring a mature AI coding product, an experienced team, and an existing enterprise customer base.

But the acquisition stalled. According to The Wall Street Journal, the issue was Microsoft, OpenAI’s largest partner. Microsoft wanted access to Windsurf’s intellectual property. Since 2021, Microsoft had used OpenAI’s models to support GitHub Copilot, which became a highlight in Microsoft’s earnings calls. But as Cursor, Windsurf, and Claude Code introduced new AI coding experiences, GitHub Copilot seemed stuck in the previous generation. If OpenAI launched a new coding product, it might not be good news for Microsoft.

This negotiation coincided with a tense period in OpenAI-Microsoft relations, as they renegotiated their partnership agreement, with OpenAI trying to weaken Microsoft’s control over its AI products and compute resources. Ultimately, the Windsurf deal was sacrificed in this game. By July, OpenAI abandoned the deal. Subsequently, Google hired Windsurf’s founding team, while the remaining staff were acquired by another AI startup, Cognition.

“I certainly hoped the deal would go through,” Altman said. “But not every deal is within your control.” He added that while he hoped Windsurf’s acquisition “could accelerate our progress,” he was also impressed by the momentum of the Codex team. During negotiations, Sottiaux and Embiricos continued developing and releasing updates.

By August, Altman decided to accelerate the overall push.

Alexander Embiricos, head of OpenAI Codex. Photo: Mark Jayson Quines.

Greg Brockman’s favorite way to measure AI capability is a small game he designed called “Reverse Turing Test.” A few years ago, he personally coded this game, and now he hands the task over to AI agents to re-implement from scratch.

The rules are simple: two human players sit at separate computers, each seeing two chat windows. One window connects to the other human, the other to AI. Players must guess which window is AI, while trying to fool their opponent into thinking they are the AI.

Brockman said that for most of last year, OpenAI’s strongest models took hours to build such a game, requiring extensive human instructions and assistance. But by December, Codex could generate a fully functional version directly from a carefully crafted prompt, powered by the new GPT-5.2 model.

This change did not go unnoticed. Developers worldwide began realizing that AI coding agents had suddenly made a significant leap. Discussions around AI coding initially focused on Claude Code, but quickly broke into mainstream media and Silicon Valley tech circles.

Even ordinary users with no coding experience started using AI to create their own software projects directly.

This surge in usage was no coincidence. During this period, both Anthropic and OpenAI invested heavily to attract more AI coding agent users. Several developers told WIRED that their $200 monthly subscriptions to Codex or Claude Code actually provided usage credits worth over $1,000. This generous quota was a market strategy: first, to get developers accustomed to using AI coding tools in daily work; then, to charge enterprise customers based on usage.

Multiple insiders said that by September 2025, Codex’s usage was only about 5% of Claude Code’s. But by January 2026, Codex’s user base had grown to roughly 40% of Claude Code’s.

A developer named George Pickett, who has worked at tech startups for ten years, recently started organizing offline meetups focused on Codex.

“I think it’s obvious we’re replacing white-collar work with AI agents,” Pickett said. “As for what that means for society, honestly, no one really knows. It will definitely cause huge disruption, but I remain optimistic about the overall future.”

Meanwhile, the productivity software company Notion, valued at around $11 billion, co-founder Simon Last said that after GPT-5.2’s release, he and his core engineering team shifted to using Codex, mainly because of its greater stability.

“I find that Claude Code often ‘lies’ to me,” Last said. “It says a task is running, but it’s not.”

Katy Shi, OpenAI researcher. Photo: Mark Jayson Quines.

Katy Shi, responsible for studying Codex behavior at OpenAI, said that although some describe Codex’s default style as “dry bread,” more users are beginning to appreciate this straightforward communication style. “Much engineering work is fundamentally about accepting critical feedback without taking offense,” she said.

Meanwhile, some large companies have already adopted Codex. OpenAI’s head of applications, Fidji Simo, said: “ChatGPT has become synonymous with AI, giving us a huge advantage in the B2B market. Companies prefer to deploy tools their employees are already familiar with.” She added that OpenAI’s core strategy for selling Codex is to bundle it with ChatGPT and other OpenAI products.

Cisco President and Chief Product Officer Jeetu Patel told employees not to worry about the costs of using Codex, emphasizing that the key is to get familiar with the tool quickly. When employees asked whether using these tools might threaten their jobs, Patel replied, “No. But I can guarantee that if you don’t use them, you will be unemployed because you will become uncompetitive.”

Today, anxiety over AI coding agents has far exceeded Silicon Valley’s tech circles. Last month, The Wall Street Journal attributed a $1 trillion sell-off in tech stocks partly to Claude Code, citing fears that software development could soon be massively replaced by AI. Weeks later, after Anthropic announced Claude Code could be used to upgrade legacy COBOL systems (still common on IBM machines), IBM’s stock experienced its worst day in 25 years.

Meanwhile, OpenAI is actively pushing AI coding agents into public discussion. The company spent millions on Super Bowl ads promoting OpenAI Codex instead of ChatGPT.

At OpenAI’s headquarters in Mission Bay, almost no one needs convincing to use Codex. Many engineers I spoke with said they rarely write code themselves anymore; most of the time, they just talk to Codex. Sometimes, they even “collaborate” collectively.

At headquarters, I attended a Codex hackathon. About 100 engineers crowded into a large room, each given four hours to produce the best demo project with Codex. An OpenAI executive stood at the front, looking at his laptop and announcing team names via microphone. Teams nervously stepped up, introducing their AI projects in slightly trembling voices. The winning team received Patagonia backpacks as prizes.

Many projects were developed with Codex, and aimed to help engineers better utilize it. For example, one team built a tool that automatically summarizes Slack messages into weekly reports; another created an internal AI guide similar to Wikipedia, explaining OpenAI’s internal services. Previously, such prototypes would take days or weeks; now, an afternoon sufficed.

On my way out, I met Kevin Weil, a former Instagram executive now heading OpenAI’s “OpenAI for Science” division. He told me Codex was helping him work overnight on some projects, and he would review results the next morning. This working style has become routine for him and hundreds of other OpenAI employees. One of OpenAI’s goals for 2026 is to develop an “automated intern” for research AI itself.

Simo said that in the future, Codex will not just be for programming but will serve as the task execution engine across ChatGPT and all OpenAI products, helping users complete various real-world tasks. Altman also expressed a strong desire to launch a general-purpose version of Codex but remains concerned about safety risks.

He said that at the end of January 2026, a friend with no technical background asked him to help install a popular AI coding agent called OpenClaw. Altman declined, thinking “it’s clearly not a good idea right now,” citing risks like accidental deletion of important files.

Ironically, a few weeks later, OpenAI announced it had hired the developer behind OpenClaw.

Many developers told me that the competition between Codex and Claude Code has never been fiercer. But as these tools’ capabilities continue to grow and are increasingly integrated into enterprise workflows, society faces issues far beyond “which AI coding tool to use.”