Dario Amodei's '90% of code written by AI in 6 months' claim revisited — and GPT-5 B2B adoption is on a tear
Sep 12, 2025
Key Points
- Coinbase reports 40% of its code is now AI-written, yet software engineering employment remains flat, suggesting companies are expanding scope rather than cutting headcount.
- Engineers using AI to write 90% of code are shipping faster and tackling harder problems, but external user-facing software quality and design remain unchanged.
- OpenAI is growing B2B adoption of GPT-5 faster than competitors, with tech and manufacturing leading, as businesses rewrite legacy systems at scale.
Summary
Dario Amodei predicted that AI would write 90% of code within six months. Software engineering employment has remained largely flat, with only modest weakness among new graduates. The claim glosses over what is actually happening inside large companies.
Brian Armstrong reports that 40% of code at Coinbase is now AI-written. Palantir and Klarna show the same pattern: flat or declining headcount paired with rising revenue. AI appears to be enabling productivity gains, though it remains difficult to separate genuine efficiency improvements from CEO directives to expand output without adding staff.
Engineers and CTOs reporting that 90% of their code is AI-generated are not working shorter days. They are writing more code, tackling harder problems, and shipping faster. The external user experience has not dramatically improved. United Airlines apps and websites across the internet have not been mysteriously rewritten or suddenly debugged. Software remains incremental.
AI lowers the friction cost of code generation, but companies are using that gain to expand scope rather than shrink headcount. More code gets written. More features ship. The quality bar and user-facing impact stay constrained by factors AI has not solved: design, product judgment, infrastructure debt, and testing complexity.
OpenAI is growing B2B usage faster than any other model vendor tracked by Ramp's AI Index. Tech and manufacturing are leading adoption. The "90% of code by AI" narrative, while technically wrong on displacement, does capture something real. Business software is being rewritten at scale, and the productivity gains are concentrated in companies with capital and engineering depth to absorb them.