On March 11, 2026, Wired published a 30-source investigation into OpenAI's effort to compete with Claude Code. The headline: "Inside OpenAI's race to catch up." A source told Wired that Codex had reached $1 billion in annualized revenue by January's end. The same day, Business Insider reported that xAI's coding agent project Macrohard had stalled amid leadership changes. The company that invented AI coding is chasing. The company that tried to enter from scratch has stopped running.
August 2021
OpenAI debuted Codex in private beta on August 11, 2021 — an API for turning natural language into code. A month earlier, Microsoft and OpenAI had announced GitHub Copilot, the first AI coding assistant for the mass market. Copilot was built on Codex. The name itself was a statement of ambition: Codex, as in the fundamental text. OpenAI didn't just enter the AI coding market. It created the AI coding market.
The results were immediate. Within months, GitHub reported that 35% of new code on its network was written with Copilot assistance. By 2023, Copilot had millions of users. Amazon launched CodeWhisperer in response. Google partnered with Replit to compete. Every major tech company built a Copilot rival. But the basic architecture remained what OpenAI established: an AI model embedded in your IDE, suggesting the next line of code as you type.
This was Wave 1. Code completion. Autocomplete on steroids. And OpenAI owned it.
February 2025
Anthropic released Claude 3.7 Sonnet and Claude Code on February 25, 2025. Claude Code was not an IDE plugin. It was a terminal tool — a command-line agent that could read your codebase, plan multi-file changes, execute shell commands, run your tests, and iterate on failures. It didn't suggest the next line. It built the whole feature.
The distinction matters. Wave 1 coding tools answered the question "what line comes next?" Wave 2 coding tools answer the question "how do I build this?" The shift from completion to agency requires different model capabilities entirely: reliable instruction-following across dozens of steps, honest error reporting when something goes wrong, precise tool use for file operations and shell commands, and sustained context management across sprawling codebases. These aren't incremental improvements on autocomplete. They're a different product category.
OpenAI launched its own Codex agent two months later, in May 2025. It ran in a cloud sandbox. By July, the terminal coding wars had begun in earnest — Claude Code, Gemini CLI, and Codex CLI all competing for developers' terminals. By October, The Information analyzed 300,000+ pull requests and found that Codex had caught up on raw metrics: a 74.3% code approval rate versus Claude Code's 73.7%.
But metrics on individual code changes turned out not to be what mattered.
The Numbers
By February 2026, the gap had widened, not narrowed.
Anthropic reported its total run-rate revenue had hit $14 billion, with Claude Code contributing $2.5 billion — up from $1 billion announced just three months earlier. OpenAI disclosed that weekly Codex users had more than tripled since the start of the year to 1.6 million. Growth, yes. But Anthropic was growing faster from a higher base.
The most striking number came from SemiAnalysis. In February, Doug O'Laughlin estimated that Claude Code authored 4% of all public GitHub commits — and was on track to cross 20% by the end of 2026. Not 20% of AI-assisted commits. Twenty percent of all commits. Bloomberg titled its profile of the phenomenon simply: "The Surprise Hit That Made Anthropic Into an AI Juggernaut."
Meanwhile, at xAI, Macrohard — the agent project Elon Musk named after buying a building for it on December 31, 2025 — stalled before it ever shipped. Co-founder Toby Pohlen left in February, the second co-founder departure in a week. The company reorganized into four divisions, but Macrohard's data project was paused. Tesla, not even merged with xAI, began ramping its own AI agent project, Digital Optimus. Musk told employees it was a "joint" effort. The reporting said otherwise.
What the Inventor Couldn't See
OpenAI's position in AI coding was, by any financial measure, dominant. Codex had $1 billion in revenue. GitHub Copilot — still built on OpenAI's models — had 15 million users. GPT-5.3-Codex, launched in February, was OpenAI's most capable coding model ever. The company was not failing. It was succeeding at the wrong thing.
The pattern is Clayton Christensen's, applied at AI speed. The incumbent with the most revenue from the current approach is the one least likely to lead the next. OpenAI had optimized for code completion — fast suggestions, IDE integration, cloud-sandboxed execution. Every dollar of Codex's billion-dollar revenue reinforced that architecture. Claude Code had no installed base to protect. It could start from scratch in the terminal, with full system access, designed from day one for autonomous multi-step execution.
But the deeper irony is about model capabilities, not product architecture. What makes a coding agent reliable — precise instruction-following, honest uncertainty, careful tool use, resistance to hallucinating file paths that don't exist — is almost exactly what safety research produces. Anthropic spent years on RLHF, constitutional AI, and refusal calibration. The goal was to make Claude trustworthy. The side effect was making Claude the model developers trust to autonomously modify their codebases. OpenAI quietly adopted Anthropic's "skills" mechanism in Codex. When the leader copies your architecture, you've already won the design argument.
Karpathy put it directly in February: AI coding agents had made "a huge leap forward since December, completing complex projects with minimal oversight." Bloomberg reported the consequence — "productivity panic among executives and engineers." The question was no longer whether AI could code. It was who controlled the tool that was rewriting the profession.
The Structural Answer
Three companies tried to build the AI coding agent that developers would hand their codebases to. One invented the category five years ago and has a billion dollars in revenue. One launched from a think tank focused on AI safety. One was backed by the richest person in the world. The safety lab leads. The inventor chases. The billionaire's project stalled.
The lesson is not about Anthropic's superiority or OpenAI's failure. It's about what happens when a market transitions between waves. Wave 1 of AI coding (2021-2024) rewarded speed, integration, and suggestion quality — the qualities you optimize when your product lives inside an IDE and your revenue comes from per-seat subscriptions. Wave 2 (2025-present) rewards reliability, autonomy, and trust — the qualities you develop when your research program is built around making AI systems that don't lie to you.
OpenAI built the best autocomplete in history. Anthropic built the model you'd trust to run git commit on your behalf. The market moved to the trust.
The same day Wired published its investigation into OpenAI's catch-up, Anthropic announced that Claude for Excel and PowerPoint now shares full context across open files — a quiet expansion into productivity software that has nothing to do with coding. And Anthropic debuted the Anthropic Institute, an internal think tank led by co-founder Jack Clark, consolidating its policy research amid an ongoing conflict with the Pentagon. The company that accidentally won the coding wars is already looking past them.
OpenAI's Codex will reach $2 billion in revenue. Maybe $5 billion. The product works and developers use it. But the Wired headline will stand as the marker: the company that created AI coding now describes its position as a race to catch up. Revenue isn't the same as leadership. A billion dollars doesn't buy you the lead when the market you invented has moved to a wave you didn't design for.