OpenAI's new reasoning models o1 and o3 are designed to excel at complex coding tasks, yet Anthropic's Claude Code has already captured nearly a fifth of the company's business, generating over $2.5 billion in annualized revenue. Meanwhile, OpenAI's Codex brought in just over $1 billion in annualized revenue by January, according to internal sources. This gap reveals a surprising truth: being first to market with a general-purpose AI chatbot doesn't guarantee dominance in specialized domains like software development. Why Did OpenAI Lose Ground in AI-Powered Coding? OpenAI's journey in coding AI is a cautionary tale of strategic pivots and missed opportunities. The company developed Codex back in 2021, an offshoot of GPT-3 trained on billions of lines of open-source code from GitHub. Greg Brockman, OpenAI's president and cofounder, demonstrated how the tool could take English commands and output code snippets, calling it a system that could "carry out commands" on a user's behalf. Microsoft quickly licensed Codex to power GitHub Copilot, which launched publicly in June 2022 and attracted hundreds of thousands of users within months. But OpenAI made a critical decision: it disbanded its dedicated Codex team. Engineers were reassigned to other projects, including DALL-E 2 (the image generator) and GPT-4 training. When ChatGPT launched in November 2022 and gained over 100 million users in two months, every other project ground to a halt. For years afterward, OpenAI had no dedicated team working on an AI coding product. The company assumed GitHub Copilot had the coding market covered and that multimodal AI (understanding text, images, video, and audio) represented the future of AI development. Anthropic took a different path. The company recognized the promise of coding earlier and trained Claude models not only on difficult coding problems from academic competitions but also on real-world, messy code repositories. When Claude Sonnet 3.5 launched in June 2024, users were impressed with its coding abilities. A startup called Cursor, which lets developers code with AI by asking for changes in plain English, saw its usage rocket upward after incorporating Anthropic's model. How Do o1, o3, and Claude Compare in Real-World Tasks? To understand where these models actually stand, consider how they perform on practical tasks that matter to users. Researchers tested flagship models from OpenAI and Anthropic across four real-world scenarios: email refinement, code debugging, structured reasoning, and strict instruction following. For email refinement, both ChatGPT and Claude produced professional, clear responses. However, Claude offered an additional feature: the ability to send the refined email directly through an integrated mail client, powered by Model Context Protocol (MCP) integration. In code debugging, the differences became more pronounced. When asked to identify a bug in a Python script that was summing array indices instead of values, ChatGPT provided an exhaustive, verbose explanation that consumed significant token space. Claude delivered a concise, direct answer with specific code recommendations that experienced programmers would appreciate. For structured reasoning tasks involving subscription tier analysis, ChatGPT again produced unnecessarily lengthy explanations. Claude not only answered concisely but added visual illustrations within the response, making complex data easier to understand at a glance. When given strict constraints, like writing a 120-word product announcement with specific formatting requirements, ChatGPT produced a dense paragraph without emphasis or formatting. Claude delivered a well-formatted response with clear visual hierarchy, making important information immediately apparent. What Makes o1 and o3 Different From Previous Models? OpenAI's reasoning models represent a fundamental shift in how AI approaches complex problems. Unlike standard language models that generate answers token-by-token, o1 and o3 work through problems step-by-step before delivering a final answer. At launch, OpenAI said o1 "excels at accurately generating and debugging complex code". Andrey Mishchenko, OpenAI's research lead for Codex, explained why this matters: "A key reason AI models have become better at coding is because it's a verifiable task. Code either runs or it doesn't, which gives the model a clear signal when it gets something wrong." OpenAI used this feedback loop to train o1 on increasingly difficult coding problems. This approach differs from how Anthropic trained Claude. Rather than focusing solely on academic coding problems, Anthropic emphasized real-world code repositories with all their complexity and messiness. Greg Brockman acknowledged this difference, noting that Anthropic's early focus on coding from messy repositories "was a lesson that we were delayed on". Steps to Evaluate AI Coding Tools for Your Workflow - Test on Your Actual Code: Don't rely on marketing claims; run both ChatGPT and Claude on real debugging tasks from your codebase to see which handles your specific coding patterns better. - Measure Response Conciseness: For experienced developers, evaluate whether the model delivers direct answers or verbose explanations that waste tokens and reading time. - Check Integration Capabilities: Verify whether the model integrates with your existing tools, like email clients or code editors, through APIs or native plugins. - Compare Cost Per Task: Calculate the actual cost of using each model on your typical coding workload, considering both token pricing and the number of tokens each model consumes for similar tasks. - Assess Instruction Following: Test how well each model adheres to specific constraints, like output length or formatting requirements, which matter for automated workflows. What Does This Mean for the Future of AI Coding? Sam Altman, OpenAI's CEO, acknowledged the competitive pressure in a recent interview. When asked why OpenAI doesn't seem to be leading the AI coding revolution, he responded: "First to market is worth a lot. We had that with ChatGPT." But he emphasized that the time is right for OpenAI to lean into coding, arguing that "it's going to be a huge business, just the economic value of it". Altman went further, suggesting that coding could be "probably the most likely path" to building artificial general intelligence (AGI), defined as an AI system that can outperform humans at most economically valuable work. He described the market as "one of these rare multitrillion-dollar markets". Yet the reality inside OpenAI has been messier than public statements suggest. The company spent much of 2023 and 2024 investing in multimodal AI models and agents, believing that the future required AI systems that could see, hear, and interact with the digital world like humans. Meanwhile, Anthropic quietly built Claude into a coding powerhouse. When OpenAI approached Cursor, the AI coding startup, about an acquisition, the founders declined, seeing the potential of the independent coding market. The gap between o1, o3, and Claude Code reflects a broader lesson in AI development: specialized focus beats generalist ambition in emerging markets. OpenAI's reasoning models may eventually catch up in raw capability, but Anthropic has already established market dominance, developer trust, and revenue streams that will be difficult to displace. The question now is whether o1 and o3's step-by-step reasoning approach can overcome that head start.
