In October, AI costs per unit of output fell for the first time in six months, by nearly 28 percent. More AI code is also being retained. Nevertheless, the bottleneck is shifting toward review, planning, and documentation.
This is according to Atlassian-data. From April through September 2026, AI spending per unit of output increased 4.5-fold, from $8.85 to nearly $40. Atlassian measures output as merged pull requests, weighted by size and complexity. In October, that figure dropped to $29.11. Not all AI costs are recorded, so the actual amounts are likely higher.
Atlassian itself acknowledges that one month does not constitute a trend. The company suspects that organizations are keeping a closer eye on budgets. In a monthly pulse survey, 40 percent of knowledge workers said they had recently been asked to justify their AI usage, had hit a budget limit, or had been switched to a less expensive tool. In addition, better context could help AI tools arrive at the right answer in fewer steps.
More AI code
The share of AI-generated code in merged code grew from 49 percent in April to 70 percent in October. How much of that will remain is another question. To measure this, Atlassian tracks code survivorship: the portion of AI-generated code that is retained in subsequent commits. That percentage dropped from 80 percent in April to 59 percent in September. In October, it rebounded to 68 percent.
However, cheaper code does not eliminate the bottleneck. In October, the number of human PR reviews grew faster than in previous months.
Workleap found its own approach. “Review time was the first bottleneck we hit, but we were able to manage it differently. We built a classifier for the type of PR, and delegated the simple ones to AI,” says CPTO Guillaume Roy.
Innovation lags behind
Developers still spend about 58 percent of their time on new features; the rest goes toward maintenance. That ratio has barely changed in nearly two years. Work related to code, such as determining what to build and coordinating with other teams, has not become any faster with AI.
In addition, developers rated 14 aspects of their work. Planning and documentation are strongly linked to code quality, but only about 62 percent of organizations have a majority of developers who view these aspects positively. Incident handling, on the other hand, receives the highest rating.
Tip: Atlassian: AI speeds up coding, but friction remains persistent