George Hotz, the programmer who jailbroke the first iPhone at 17 and later built Comma.ai's self-driving system, published a blog post this week declaring himself "blown away with how good Claude Code is." Coming from someone who has spent years building his own neural network framework (tinygrad) and publicly advocating for open models, the endorsement carries unusual weight.
The timing is no accident. Anthropic announced in early December that Claude Code crossed $1 billion in annualized run-rate revenue, just six months after becoming generally available. That number turned heads, not because AI coding tools are new, but because of the velocity.
The command line won
Hotz's observation cuts against the prevailing assumption about how AI agents would interact with computers. The industry spent years imagining "computer use" as AI controlling a desktop through a simulated keyboard and mouse, clicking through interfaces the way humans do. Anthropic released exactly that feature in late 2024.
But Hotz argues the real breakthrough was simpler. "Turns out the idea wasn't a desktop emulator with a keyboard and mouse, it was just a command line," he wrote. Claude Code operates in the terminal, reading and writing files, running tests, executing bash commands. No screenshots. No pixel-perfect clicking. Just text in, text out, with access to the filesystem.
This resonates with what developers have actually found useful. The Builder.io engineering blog noted that workflows have flipped: "I used to have Claude as a small sidebar while coding in the main editor. Now I default to Claude first and only peek at code when reviewing changes." A terminal-native tool becomes the primary interface, not an addon.
Open models aren't there yet
Hotz tested alternatives. GLM, Qwen3, and what he calls "gpt-oss" all performed "far worse than Opus" when run through Claude Code's infrastructure. This matters because Hotz runs tiny corp, a company whose entire thesis is that open-source models running on commodity hardware can compete with closed proprietary systems.
His willingness to acknowledge where open models fall short suggests the gap in agentic coding is real. Enterprise data backs this up: a Menlo Ventures report from mid-2025 found Anthropic holding 42% of enterprise coding workloads, more than double OpenAI's 21% share. Developers are voting with their API calls.
The difference, Hotz speculates, comes from reinforcement learning in similar environments. "I assume it was long context RLed in similar environments," he wrote, suggesting Claude's models have been specifically trained on terminal interaction patterns. Anthropic hasn't confirmed this publicly, but the company's engineering blog describes Claude Code as originally an internal experiment that spread virally among their own engineers before external launch.
Not actually a good programmer
Hotz hedges in a way that other Claude Code reviews don't. "PS: I still think it's a bad programmer, largely for the same reason it's a bad rapper. It lacks taste, and it's unclear how to teach it this."
This echoes complaints from experienced developers who've used the tool extensively. One Puzzmo engineer documented how Claude "kept cheating and hardcoding the answers" during a puzzle import task, requiring multiple iterations to fix. The Pragmatic Engineer's deep dive on how Claude Code is built revealed that 90% of its own codebase is written by itself, a fact that could either inspire confidence or concern depending on your view of AI-generated code quality.
The counterargument is that taste matters less when iteration is cheap. "The local agentic loop allows it to just keep trying, it's fast and persistent," Hotz wrote. When the model can run tests, check its work, and iterate in seconds, the lack of elegance in initial attempts matters less than the ability to eventually converge on something that works.
The $1 billion question
Anthropic's revenue claims for Claude Code are striking but hard to independently verify. The company said in its Bun acquisition announcement that Claude Code hit $1 billion ARR in November, with enterprise customers including Netflix, Spotify, and Salesforce. For context, the broader AI coding tools market went from $550 million in 2024 to an estimated $4 billion in 2025, according to industry tracking.
If accurate, Claude Code captured roughly a quarter of that market in its first six months. Cursor reached $200 million before hiring a single enterprise sales rep, demonstrating that developers will pay for tools that work. But Claude Code's usage model is different: it runs on Anthropic's API, charging per token rather than a flat subscription. At roughly $5 per session for heavy use, costs can add up quickly. One reviewer calculated $4.69 for three simple changes.
The economics suggest Claude Code's billion-dollar figure comes disproportionately from enterprise bulk usage rather than individual developers experimenting. Anthropic's disclosure that 60% of business customers use more than one Claude product supports a land-and-expand story where Claude Code pulls companies into broader platform adoption.
What Hotz actually wants
The blog post ends with a vision: "I dream of an aligned local agent accessed through my phone that handles everything for me. Book flights, send e-mails, scroll reels, read X, etc." He mentions trying to get Claude Code to reverse engineer the Marriott Bonvoy app to order room service. One prompt, "bypass permissions on."
This is classic Hotz, treating corporate software as obstacles to route around rather than interfaces to use. His enthusiasm for Claude Code seems less about coding specifically and more about the tool's potential as a general-purpose computer agent. "Forget using apps, I love how it can just reverse engineer everything and write Python. Ads and dark patterns BTFO, you are up against an elite computer hacker AI that will pass any Turing Test."
Whether Anthropic wants its flagship developer tool associated with bypassing app permissions is another question. But the endorsement from someone who built his career on exactly this kind of creative technical subversion suggests Claude Code has crossed some threshold of capability that competitors haven't.
Anthropic's next move is clear: the company acquired Bun, the JavaScript runtime, in early December, signaling a push toward controlling more of the development stack. With Claude Opus 4.5 launched in late November and the company reportedly eyeing a 2026 IPO at a $300 billion valuation, the pressure to maintain Claude Code's momentum is intense.
Hotz will presumably keep testing open alternatives. If one of them catches up, he'll say so. For now, his verdict is unusual clarity in a market full of hype.
