A year ago this month, Anthropic CEO Dario Amodei told the Council on Foreign Relations that AI would be writing 90% of code within three to six months, and "essentially all of the code" within twelve months. We've hit that twelve-month mark. So how did the prediction hold up?
Not well, by industry-wide numbers. Google reported about 25% of its internal code was AI-generated as of late 2024. Microsoft said roughly 30% by mid-2025. Those are meaningful numbers, but they're a long way from "essentially all."
The Metric Problem
Amodei wasn't making the claim from thin air. By October 2025, speaking at Dreamforce alongside Salesforce CEO Marc Benioff, he said 90% of code at Anthropic itself was being written by AI. That's a bold internal stat, and researchers at Redwood Research actually investigated it. Their conclusion: the number is "probably closer to 90% than 50%" if you count any code where AI was involved at all. But that's not what most people mean when they hear "AI writes 90% of code." There's a wide gap between "AI drafted this and a human cleaned it up" and "AI independently produced working software."
At Davos in January 2026, Amodei doubled down. He told The Economist that some Anthropic engineers "don't write any code anymore" and just edit what the model produces. He placed the industry six to twelve months away from AI doing "most, maybe all" of software engineering end-to-end.
The Counterarguments
Not everyone in the industry agrees. IBM CEO Arvind Krishna publicly estimated only 20-30% of code could be written by AI, arguing the technology boosts programmer productivity rather than replacing it. Google DeepMind CEO Demis Hassabis gave only a 50% chance of achieving human-level AI within the decade. Meta's Chief AI Scientist Yann LeCun went further, arguing that large language models will never achieve human-like intelligence and that a fundamentally different approach is needed.
Developer surveys tell a more grounded story: 84% of developers use or plan to use AI coding tools, and 51% use them daily. But only 20% say they trust AI-generated code completely.
What's Actually Happening on the Ground
The practical reality sits somewhere between Amodei's prediction and the skeptics' pushback. AI coding tools have genuinely changed daily workflows. Anthropic's own Claude Code launched in February 2025 and went broadly available by May. The company followed up with Claude Code Review in March 2026, a multi-agent system that crawls codebases looking for bugs. Claude Cowork expanded the approach beyond developers entirely.
These tools are real and useful. But "useful" and "writes essentially all the code" are different claims. Most developers are using AI as a sophisticated autocomplete and first-draft generator, not as an autonomous software engineer. The human is still steering, reviewing, debugging, and making architectural decisions.
Amodei's prediction was probably right about the direction and wrong about the speed. AI code generation is growing fast, but the jump from "helpful assistant" to "writes all the code" requires more than just better models. It requires trust, tooling, and organizational change that moves slower than model capabilities. The 90%-at-Anthropic stat might be real, but Anthropic is a company built around these models, staffed by people eager to push the limits. Most engineering organizations aren't there, and won't be for a while.
Futurism ran a piece in September 2025 titled "Exactly Six Months Ago, the CEO of Anthropic Said That in Six Months AI Would Be Writing 90 Percent of Code." That headline did the fact-checking on its own.