OpenAI CEO Altman: To Catch Up with Claude Code, Atlas Browser Product Shut Down and Video Generation Product Sora Ceased
OpenAI CEO Sam Altman recently admitted in an interview that the company was previously "far behind" Anthropic's Claude Code in the programming agent field. The decision to allocate resources to directly compete was seen internally as "like a crazy kamikaze mission."
Altman recalled that the general consensus within the team was that "such a comeback would never succeed," but the company still deemed it crucial, mobilizing the team to tackle the challenge. Ultimately, they achieved what he described as a "rare and incredible" turnaround in business history, with Codex now being used by most top programmers he knows.
According to third-party data, in September 2025, Codex's weekly active users were only 5% of Claude Code's. After the desktop version launched in February 2026, the gap quickly narrowed, with Codex's weekly active users rising from about 3 million in early April to over 7 million by July 14, and surpassing 10 million by the end of July, bringing the total weekly active users for both products to 10 million.
In terms of revenue, Claude Code's annualized revenue surpassed $2.5 billion in February 2026, while Codex just crossed $1 billion in January 2026. To close the gap, OpenAI expanded its programming reasoning-related computing power by about 100 times from the beginning of the year to July, with Altman admitting that this required temporarily putting other product experiences like writing and conversation "on hold."
This turnaround came at a high cost—OpenAI shut down multiple business lines to concentrate resources: the video generation product Sora was ceased in March, despite a $1 billion collaboration with Disney in progress at that time; the independent browser product Atlas was cut just 9 months after its launch; meanwhile, Codex offered nearly unlimited quota resets to users, averaging a reset every 9 days in some months, and even every 3 days in July, with the actual model cost per user often exceeding $1,000, far above the $200 subscription price.
In market mechanics, Altman views programming agents as a key entry point to "recursive self-improvement" (models participating in their own training iterations), thus determining that they "cannot afford to lose." This high-investment, high-subsidy share battle essentially involves trading short-term massive losses for developer mindshare and long-term training data feedback loops, with funds flowing back from video, browser, and other consumer scenarios to programming, seen as the core hub for AI capability self-enhancement.
Source: Public Information
ABAB AI Insight
OpenAI previously established a leading position in ChatGPT through "first-mover advantage," but in the niche of programming agents, Anthropic, with Claude Code, took the lead in developer mindshare. This is similar to OpenAI's historical pattern of catching up in image generation and voice assistant fields—demonstrating its ability to achieve short-term turnarounds by concentrating resources in lagging situations.
To catch up with Claude Code, OpenAI has shut down non-core products like Sora and Atlas to free up computing power and funds, while also providing Codex users with nearly unlimited quota subsidies at prices far below actual costs. Essentially, this reallocates capital and computing power from multiple product lines to focus on programming as a single battlefield, trading short-term cash burn for market share and developer data.
This strategy aligns with the early "burning money for market" subsidy wars of ride-hailing and food delivery platforms—using pricing far below costs to secure user habits and data barriers, then gradually narrowing subsidies once the market stabilizes. The difference is that the moat for programming agents is not only user scale but also whether the model can leverage real programming data to enhance its own training. This places the current AI programming tools sector in a critical window of "data flywheel competition."
At its core, this is a technological substitution—OpenAI's bet on programming agents is not just to capture developer tool market share but also to view it as a key infrastructure for achieving "recursive self-improvement." Once models can participate in their own capability iterations at scale through programming agents, it will fundamentally replace the current model development paradigm that relies on manual labeling and design training processes. This is also the deeper logic behind Altman's willingness to incur significant short-term losses for this endeavor.
ABAB News · Cognitive Law
A battlefield that cannot afford to lose is worth fighting for with everything.
The subsidies burned are not just money; they are tickets to the next generation of models.
Stories of turning the tide against the wind always begin with a consensus that it is impossible.