Claude Code Creator: 100% AI Code Is Now "Problematic"

Boris Cherny is walking back his own hype. The Anthropic engineer who declared in January(opens in new tab) that AI had "solved coding" — and who personally writes zero lines of code — now admits that companies running 100% AI-generated code are hitting a wall. The admission matters because Cherny is not a critic. He runs Claude Code. When the person building the tool says the tool creates problems, pay attention.

What happened: Cherny walks back "AI solved coding"

Speaking at a Scale AI fireside chat in June 2026, Cherny was asked directly about Uber COO Andrew Macdonald's concerns that AI spending wasn't producing enough ROI. Macdonald had reason to worry — Uber burned through its entire 2026 AI budget in four months (Business Insider, 2026).

Cherny's response was a pivot from his January posture (Times of India, 2026):

"ROI is absolutely the right framing because you don't want to just think about cost — you kind of spend something on it and you get something back."

He then outlined an implicit three-phase maturity model for AI coding adoption:

Cherny now calls Phase 2 "problematic." When every engineer's throughput skyrockets, the bottleneck shifts from code production to code review, quality assurance, and — critically — having enough good ideas to feed the machine.

He acknowledged that Anthropic itself feels the cost pressure: "Every token we use is a token we do not give to a customer" (Business Insider, 2026). That's a telling admission from a company that makes money selling tokens.

Why it matters

Cherny's pivot is not a personal correction. It's an industry signal. The AI coding tool wave is entering a realism phase, and the conversation is shifting from "look what it can do" to "here's what it costs and here's what you get back."

OpenAI's Sam Altman is making similar noises about enterprise ROI concerns. The pattern is clear: the companies that built the tools are now the ones warning about the tools' limits. That's worth more attention than any third-party criticism.

For Claude Code(opens in new tab) specifically, this reinforces the verdict we already hold: conditional. The tool is architecturally ahead — dynamic workflows, maker-checker sub-agents, cloud routines — but the cost-control infrastructure that separates a powerful demo from a sustainable enterprise tool is still catching up. Cherny himself says companies should give employees "tokens and safety to experiment" while controlling costs "on the backend, not on the front end." It's a difficult balance that no tool has fully solved.

What changes for you

If you're evaluating AI coding tools in mid-2026, the capability benchmark is no longer the right starting point. Cherny's own words frame the new priorities:

• Budget for tokens, not just seats. Autonomous agent loops compound costs rapidly. If your tool doesn't give you hard spending caps, an unattended loop can burn a month's budget in a weekend.
• Measure ROI, not adoption percentage. "What percentage of code is AI-generated?" is a vanity metric. "What did we ship faster, and at what cost?" is a business metric.
• Solve the idea bottleneck early. Cherny's Phase 3 warning is underappreciated. If your engineers can produce 10x more code but your product org can only define 2x more features, you've bought a faster treadmill, not a faster company.

FAQ

Is 100% AI-generated code really a problem?

Cherny himself says yes. When every engineer's throughput skyrockets, the bottleneck shifts from code production to code review, quality assurance, and — critically — having enough good ideas to feed the machine. The code gets written faster than humans can review it.

What should companies measure instead of AI adoption percentage?

ROI — features shipped per dollar of AI spend. Cherny agrees: "ROI is absolutely the right framing." Measure what you ship faster and at what cost, not what percentage of your code an AI wrote.

Is Claude Code still recommended?

Neomanex holds a conditional verdict on Claude Code. The tool is architecturally ahead — dynamic workflows, maker-checker sub-agents, cloud routines — but cost-control infrastructure is still catching up. It's the right tool for teams that can manage token budgets and have a strong review culture. It's the wrong tool for teams that want to set it and forget it.