I've been writing software for the best part of 30 years at this stage. I've directly employed about 50 software engineers, and worked with many hundreds more around the world, from the largest companies to the smallest.
I have a reasonably good feel for the craft, and for the effort it takes to ship production grade / business critical products to demanding customers.
The productivity improvement with frontier models can be very large, but it is not one number. In my experience, you see everything from marginal gains to order-of-magnitude gains depending on task type, context quality, and review discipline.
If you force me to put rough numbers on it: 2-5x is common for well-scoped implementation work, 5-20x appears in strong workflows, and extreme cases can approach 50x for narrow tasks with excellent context and tooling.
The move from "interesting autocomplete" to "serious production accelerator" happened quickly. Reliability is still uneven, but the gap can be managed with guardrails:
- clear specs and constraints up front
- deterministic checks in CI
- human triage on high-impact changes
- short review/fix loops
What does this mean when a team of 10 can out-ship much larger teams from a few years ago?
(1) Large companies will struggle to transition quickly
Adoption will be uneven. Some teams have already redesigned how they scope, review, and ship. Others are still testing tools at the edge of their workflow.
That gap creates a temporary advantage for small, fast teams in markets where incumbents are slow to change operating model.
(2) There are a lot of problems to solve in the world
If your effective engineering capacity increases, a lot of previously uneconomic problems become viable. Legacy maintenance, internal tooling, and niche vertical software all get more attractive.
But leverage is not automatic. Without quality gates, teams simply ship bugs faster.
The upside is real. The execution burden just moved from typing code to designing better systems for generating, reviewing, and operating it.
For my own workflow, model releases around August 2025 were the turning point where this went from novelty to default.