Eighteen months. That's how much closer the "AI replaces coders" timeline just moved in a single quarterly update.
The AI Futures Project, a forecasting group led by former OpenAI researcher Daniel Kokotajlo, published their Q1 2026 timelines revision this week. The headline number: their median prediction for when AI can fully automate software engineering shifted from late 2029 to mid-2028. Co-forecaster Eli Lifland made a similar shift, moving his estimate from early 2032 to mid-2030.
The milestone they're tracking is called "Automated Coder" (AC), defined with brutal clarity: the point when an AI company would rather fire all its human software engineers than stop using AI for coding. It's a deliberately extreme definition, and the fact that serious researchers think it's getting closer faster should get your attention.
What Changed in Three Months
Three things drove the update. First, the team switched to a newer version of METR's coding time horizon benchmark (v1.1), which measures how long and complex a coding task AI can handle autonomously. Second, they evaluated recent models including Gemini 3, GPT-5.2, and Claude Opus 4.6, and found the capability doubling time (how fast AI coding ability doubles) accelerated from every 5.5 months to every 4-4.5 months.
Third, real-world adoption data. The team specifically cited Claude Code reaching an annualized revenue of over $2.5 billion in early February 2026, just nine months after launch. When a single AI coding tool generates that kind of revenue that fast, it shifts the baseline for what "adoption" looks like.
How Seriously Should You Take This?
These forecasts come with massive uncertainty bands. Kokotajlo's 80% confidence interval still spans several years, and Lifland remains more conservative by about two years on most milestones. Forecasting AI progress has a poor track record in both directions: people overestimated self-driving cars and underestimated large language models.
But the direction matters more than the exact date. Every quarterly update from this group has moved the timeline forward, not back. The METR benchmark trend line they rely on, which tracks AI performance on increasingly complex coding tasks, has been remarkably consistent. And the revenue numbers suggest businesses aren't waiting for the forecasters.
For anyone building a career around software engineering, the practical takeaway isn't "learn to code or don't learn to code." It's that the tools sitting alongside you are getting better at a compounding rate that keeps surprising even the people whose full-time job is predicting this stuff. Whether the AC milestone hits in 2028 or 2032, the years between now and then will see steady, significant changes in what AI handles vs. what humans handle in a typical engineering workflow.
The forecasters plan quarterly updates. Based on the trend, don't expect the next one to push dates further out.