Writing code was the killer app that proved AI could do real work - not just chat. Two years ago, that meant GitHub Copilot autocompleting your functions. Today, it means Claude Code autonomously editing files, running tests, and pushing commits while you watch. The distance between those two things is the story of the AI coding wars.
From Autocomplete to Agents
The first wave of AI coding tools were essentially fancy autocomplete. GitHub Copilot, launched in 2021, suggested lines and blocks of code based on context. Useful, but still fundamentally you doing the work with a smarter assistant.
The second wave changed the model entirely. Tools like Claude Code and Cursor operate as agents - they can read your entire codebase, run terminal commands, fix failing tests, and work through multi-step problems without constant hand-holding. You describe the bug; they find it and fix it.
Then came "vibe coding" - a term coined by researcher Andrej Karpathy in early 2025 for the practice of describing what you want in plain language and letting AI write all the actual code. Non-developers building functional apps. Startups prototyping in hours instead of weeks. The skill being tested shifted from whether you can write a function to whether you can describe what you want clearly enough.
Three Companies Targeting the Same Person
OpenAI, Google, and Anthropic are all chasing the same user: the developer (or aspiring developer) who wants AI to handle more of the actual coding work.
OpenAI has ChatGPT for conversational coding plus deep integrations with third-party editors. Google is pushing Gemini Code Assist at enterprise development teams. Anthropic built Claude Code as a terminal-native agent - something you run in your command line rather than inside a browser tab, which means it fits into existing workflows without forcing a tool switch.
The benchmarks tell part of the story. SWE-bench is a standardized test that measures how often an AI can independently solve real GitHub issues - actual bugs from real open-source projects, not toy exercises. Scores have jumped from around 1-3% in early 2024 to over 50% for top models today. That's a real capability shift, not just better marketing.
What Experience Actually Shows
Benchmarks don't tell you what it's like to use these tools on a real codebase at midnight with a deadline.
From actual use: context window size (how much code the AI can hold in working memory at once - think of it as how many pages of a codebase it can read before forgetting the beginning) matters more than raw benchmark scores for anything beyond toy projects. Speed matters more than it should - a 30-second wait for a code suggestion breaks your thinking rhythm. First-attempt accuracy matters most of all, because correcting three wrong answers from an AI takes longer than writing the code yourself.
All three competitors are improving on all three dimensions, and meaningfully. Tools available now are faster, more accurate, and handle larger codebases than anything from 12 months ago.
The winner will probably be whoever gets the IDE and terminal integration right. The best AI coding tool is the one already in your workflow - not the one you have to context-switch to use.