Amodei's 90% AI code prediction: a year on, what stuck

In March 2025, Dario Amodei, Anthropic's CEO, sat across from Mike Froman, president of the Council on Foreign Relations, and dropped a sentence that silenced the room. Fifteen months later, in mid-2026, we have enough perspective to look it in the eye. This article does exactly that work: restore what was said precisely, measure what came true, name what didn't happen, and draw the practical consequence for those who create.

"I think we'll be there in three to six months, where AI is writing 90 percent of the code. And then in twelve months, we may be in a world where AI is writing essentially all of the code."

Dario Amodei, Council on Foreign Relations, March 10, 2025

The sentence got picked up everywhere, often shortened to "90% of the code in a year." That shortcut fueled a polarized debate between skeptics who saw a marketing stunt and enthusiasts who announced the end of the developer trade. The truth, as so often, requires going back to the text.

What Amodei actually said

The exact passage runs in two beats. First, the volume prediction: three to six months for AI to write 90% of the code, twelve months to approach the totality. Then, and this is the part almost systematically cut out, a clarification on what the developer still owns: "The programmer still needs to specify what are the conditions of what you're doing, what is the overall app you're trying to make, what's the overall design decision."

In other words, from the original wording, Amodei did not say developers would vanish. He said the share of their work spent typing code would flip to the machine, while the share spent deciding, specifying, judging and orchestrating would stay on the human side. That nuance didn't survive the news cycle. Yet it's the whole point.

To place the prediction, you need to recall the moment. March 2025 came three months after the public release of agents that actually execute code in loops, and two months after the first serious agentic coding tools appeared at the major editors. Amodei wasn't projecting into a void. He was extrapolating a curve he was already watching internally at Anthropic.

What actually came true

Fifteen months later, let's look at the ground without complacency.

The uptake of AI-assisted coding tools has been massive. Claude Code at Anthropic, Codex at OpenAI, Cursor, Windsurf, GitHub Copilot in its agentic form: these tools left the early-adopter zone to enter the everyday practice of development teams. A third-party survey from late 2025 estimated more than 70% of professional developers use an AI assistant daily, against less than 30% when Amodei made his prediction. The curve is clear.

Inside the companies that pushed furthest, the share of AI-generated code became measurable. Several analysts and internal accounts place that share around 70 to 90% in teams that truly embraced agentic coding. At Anthropic itself, public analyses estimate the fraction of useful code written by AI is closer to 90% than 50%, even if the measurement remains hard to frame rigorously.

On that point, the prediction holds. Not literally within six months for the whole industry, but with a propagation speed that justifies Amodei's stance. Software development transformed faster than most competing predictions thought possible.

What did NOT happen

Developers didn't disappear. By mid-2026, employment for senior developers and architects remains tight, sometimes tighter than before. Pure juniors, on the other hand, are going through a hard zone: companies figured out that a developer orchestrating Claude Code easily replaces three or four juniors typing by hand. But the trade itself didn't implode.

The work shifted. Where a developer used to spend 70% of the time writing code and 30% designing and debugging, you now see the reverse: 20% typing, 80% specifying, reviewing, making architectural decisions, orchestrating agents. The rare skill is no longer typing velocity. It's the quality of judgment and the finesse of specification.

💡 Amodei's prediction didn't kill the developer trade. It killed the part of the trade no serious developer ever loved: repetitive typing, boilerplate, plumbing. What remains is precisely what makes the difference.

Another thing that didn't happen: code quality didn't collapse. Initial worries about the proliferation of AI-generated bugs proved overblown once teams put a review discipline in place. Automated review functions, LLM-style linters, and extended testing practices absorbed the new velocity. What did drop, on the other hand, is the patience for poorly architected code: a team that produces ten times faster has ten times fewer excuses for shipping technical debt.

The real thesis behind the sentence

Re-reading Amodei's statement today, with fifteen months of perspective, you understand that the sentence wasn't a percentage forecast. It was the announcement of a shift in the very nature of the trade.

Before, the developer was a craftsman-typist. The cardinal skill was translating an intention into lines of code, and that translation took time. The value sat in the speed and accuracy of the translation. Now, the translation became a commodity. What the machine does for free and instantly, you no longer pay humans for.

The cardinal skill shifted. It's no longer in the typing, but in three disciplines that, oddly, resemble the work of old-school project leads: knowing exactly what you want, knowing how to phrase it precisely enough for a non-human intelligence to execute it, knowing how to judge what comes out. The developers who master those three gestures see their value climb. Those who stay anchored on pure typing skill see their position erode.

Dimension	Old developer (pre-2025)	New developer (post-Amodei prediction)
Primary output	Lines of code typed by hand	Clear specifications, sharp reviews, agent orchestrations
Rare skill	Typing velocity and syntactic mastery	Architectural judgment and quality of specification
Added value	Translating intention into functional code	Deciding what must exist, judging what comes out, guaranteeing coherence
Posture	Patient executing craftsman	Conductor of AI agents, guardian of overall quality

This shift isn't exclusive to developers. It hits every trade where part of the work consists of producing a structured deliverable from a fuzzy intention: writers, designers, analysts, project managers, product owners. Wherever you translate intention into artifact, the machine invites itself. And everywhere, the same question comes up: which share of my work was typing, and which share will remain decision?

What this changes for creators, not just developers

For a creator who has never written a line of code, Amodei's prediction can feel distant. It isn't. The same motion that shook software development is now reaching image, video, text, presentation, and business plan production. The gesture of directing an AI agent to produce something usable became the number one transversal skill.

Learning to direct agents is no longer optional. It's, by mid-2026, the condition to stay relevant in a market where the machine handles the typing and human value concentrates on intention, taste, and judgment. As we wrote in our article on Claude and creativity, the tool without the hand that knows what to do with it stays a sleeping piano.

And as we wrote in our piece on the AB-Arts method, that hand isn't acquired by reading the docs. It's built through curiosity, interest, and patient documentation of what you've tried. Amodei's prediction doesn't invalidate that interior work. On the contrary, it makes it more urgent.

The AB-Arts position

We read Amodei's prediction as good news, provided we welcome it lucidly. The good news is that the least interesting part of every creative trade flips to the machine. Nobody loved spending three hours debugging a missing character, reformatting a table, redoing the same paragraph five times to adjust the tone. That work, now, the agent does. The time you reclaim, you can put elsewhere.

The lucid condition is that you have to learn to hold the instrument. Directing a Claude agent so it writes useful code, produces a relevant report, drafts a video worth watching, isn't a skill you improvise. It builds through framed practice, in a setting that helps structure the learning.

💡 Human creativity isn't replaced by AI. It's multiplied by those who know how to direct it. The Claude masterclass exists precisely to transmit that gesture.

That's exactly the reason for our Claude masterclass. We train professionals (developers, creators, executives, designers) to direct Claude agents so they serve their real work, not to run demos full of empty promises. The training happens in small groups, in Brussels, on participants' real projects. By the end, we don't form chatbot users. We form people who hold the instrument with precision.

If Amodei's prediction concerns you, if you sense your trade is shifting without quite knowing how to position yourself, the Claude masterclass is built for that. For a whole-team engagement in your studio or organization's context, reach out through our contact page. We'll discuss a program tailored to your concrete stakes.

---

Amodei's sentence is fifteen months old. Developers are still here. But their trade flipped, quietly and deeply, toward what it should always have been: a work of decision and judgment, supported by an intelligence that does the rest. It's up to each of us to decide whether to stay on the typing shore, or learn to drive the new vessel. To dig further on this same reflection, read our article on Claude as an instrument and our method to stay curious before automating.