Habr AI: METR and Google Cloud Find No Promised Speedup for Developers Using AI

Habr AI has published a review of several studies on how AI tools affect programmers' work. The main conclusion is an uncomfortable one for the market: developers often feel they are working faster, but measurable indicators do not yet confirm the promised speedup.

Where growth is not visible

One of the key arguments in the article concerns the Python ecosystem. The author cites an Answer.AI analysis that examined PyPI dynamics in search of traces of "explosive" productivity. If AI really sped up development radically, it should show up in the number of new packages, release frequency and the overall volume of published code. But the picture turned out to be far more modest: there is no noticeable jump in new packages, and the rise in update frequency began back in 2019 and is probably linked more to CI/CD practices than to generative models.

An interesting detail is that the increase in activity is visible mainly in AI projects. Such packages are updated more often, but the author interprets this not as a universal effect of the tools but as a consequence of hype and an influx of investment into that segment. Similar logic applies to GitHub data: according to the review, there is no mass surge of new repositories either, although that would be a simple indicator that launching side projects has become significantly easier. The conclusion is blunt: AI helps build impressive prototypes but does not remove the real bottlenecks in development and product launch.

What the studies show

The most telling gap is between how teams feel and what operational metrics show. The article cites a Google Cloud report on the impact of GenAI on software development: 75% of developers say AI makes them feel more productive. But at a 25% adoption rate of such tools, delivery throughput falls by 1.5% and delivery stability by 7.2%. In other words, working with AI is subjectively more pleasant, but that does not necessarily mean more useful output.

• 75% of developers feel a productivity gain
• delivery throughput falls by 1.5%
• delivery stability drops by 7.2%
• the surge in activity is most visible in the AI segment

The METR study the author cites looks even harsher. In it, experienced developers expected AI tools such as Cursor and Claude to speed up their work by about 20%. In fact the opposite happened: tasks took about 20% longer. The explanation seems plausible: engineers type less code by hand but spend more time on review, fixes, waiting for the model's response and re-running. Some of the mechanical routine goes away, but a new layer of quality control appears in its place.

"Today we have evidence that AI makes many of us not more, but less productive."

This gap between expectation and reality matters not only for engineers but also for business. The piece also gives the example of Notion: after adding AI features, the product's margin, according to the company's CEO, fell from 90% to 80%. The logic is clear: the market pushes companies to build AI into almost everything, but extra spending on inference and infrastructure guarantees neither audience growth nor revenue growth. For companies, this means that adopting AI in development and in products cannot be judged by the wow effect; it takes metrics for speed, stability and economics.

What this means

The Habr AI review cools expectations around "tenfold" speedups in programming. At this stage, AI coding improves the work experience and speeds up individual parts of the process rather than delivering a guaranteed gain in team or business productivity. For managers the conclusion is simple: before rolling out such tools at scale, measure not developers' mood but cycle time, delivery quality and the total cost of changes.