TL;DR
Get business pricing on networking and server gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
The Pragmatic Engineer gathered AI spending and pull-request data from 27 workplaces, alongside aggregate figures from CodeRabbit and Gitar. The figures show higher PR volume in some samples and wide variation in per-engineer costs, but the report says few companies can yet establish whether the spending delivers a measurable return.
The Pragmatic Engineer’s report on AI spending at 27 workplaces found that companies are paying widely different amounts for coding tools, while some data sets show engineers opening more pull requests than a year earlier. The report does not establish that AI caused the increase or that it produced a financial return: its author says few companies can answer the ROI question with confidence.
The report combines interviews with engineering leaders and senior engineers at 27 workplaces with aggregate data shared by CodeRabbit and Gitar. The companies’ spending varied sharply: one two-person .NET shop reportedly paid $3,000 per engineer per month, while a publicly traded company with 1,500 engineers capped monthly spending at $150 per developer. The author found no clear link between company size and per-engineer outlay.
CodeRabbit’s anonymized platform data showed a 113% rise in weekly pull requests among surveyed data engineers and a 127% increase in monthly PRs. A separate comparison tracking the same 20,000 developers showed a 128% increase over a year. These are changes in activity recorded on the platform; the report does not show that AI tools alone drove them or that more PRs translated into better software or lower costs.
Gitar, an AI code-verification agent from Sonar, reported that some teams it categorizes as “AI-native” produce 60 or more PRs per developer per week. The report also describes spending patterns rather than a controlled productivity study: respondents said costs peaked in June and July and later fell, in some cases by roughly half for API-based plans. The author attributed that decline to a mix of model changes and use of open models, based on what companies told him.
More PRs, Unproven Returns
The findings show why companies are trying to put a number on AI use: tool bills are growing, and executives want to know whether those costs improve engineering output enough to justify them. Pull-request volume is an accessible measure that organizations can compare over time, but it captures quantity, not necessarily quality, usefulness, or business value.
The gap matters for both employers and engineers. A team that reports more PRs may still face extra review work, rework, or maintenance costs; the report does not quantify those effects. Meanwhile, monthly spending can be obscured by subsidized subscriptions, usage caps, and different pricing plans. The reported figures therefore offer a snapshot of adoption and activity, not a like-for-like calculation of return on investment.
Companies also face different exposure to costs. The report says spending caps are common among medium- and higher-spending organizations, although respondents described those limits as easy to work around. Some companies said subsidies were ending, which could raise bills substantially if usage stays constant. That makes tracking both cost per engineer and outcomes more pressing, without resolving what outcome should count as success.
Products Worth Considering
EcoFlow Delta 3 Max Portable Power Station with 220W Solar Panel, 2048Wh
Bose TV Speaker (2nd Gen) Compact Soundbar
LG UltraGear 25G590B-B 25-Inch Full HD 1000Hz Gaming Monitor
As an affiliate, we earn on qualifying purchases.
How the Spending Figures Were Gathered
The article grew from a reader’s question about AI spending per engineer and how PR throughput compares before and after adoption. Its author sought responses from 27 workplaces and added two vendors’ aggregate platform data. Most company names were withheld at respondents’ request, and the report presents a range of examples rather than a representative survey of the technology industry.
The report describes spending from low-cost plans to four-figure monthly outlays per developer. It says Claude models, including Opus 5.5 as named in the source, were frequent destinations for spending, alongside Cursor, Codex, and GitHub Copilot. The author also reported that some companies moved toward open models as costs changed. Those details reflect the companies contacted and should not be read as a market-wide ranking of tools or spending.
PR counts are useful because many companies already track them, but the report notes that team norms shape output: a team accustomed to a handful of PRs a week may work differently from one where developers routinely ship many more. The comparisons do not adjust for those differences, job mix, codebase complexity, review standards, or changes in staffing. That limits what can be inferred from the increases.
““What’s the ROI of AI? A question that few companies seem able to answer – at least for now.””
— The Pragmatic Engineer report author
What the Data Cannot Establish
The report does not provide a controlled test showing that AI tools caused the reported PR increases. It also does not compare the value or quality of the PRs, measure time saved, or calculate whether savings exceeded subscription and API costs. A direct ROI figure is absent, and the author says most companies contacted could not supply one.
The evidence has limits: company spending details came from a small set of workplaces, many anonymized, while CodeRabbit and Gitar supplied data from their own platforms. The source does not provide enough methodology to assess whether those platform samples represent software engineers generally. It is also unclear how much changes in work practices, team composition, or measurement contributed to PR counts.
Some details are time-sensitive. The report says spending peaked in June and July and that API costs later fell for some firms, but it does not give a common observation period for every company. It also describes named model versions and prices that may change. The source material’s account of one firm’s plans for internal computing infrastructure is incomplete, so its expected future spending and operating model cannot be established from the available details.
Companies Need Outcome Measures
The report points to continued efforts by companies to track AI bills and compare them with engineering output as usage expands. For a meaningful ROI assessment, organizations would need to connect tool costs with outcomes such as delivery time, review burden, defect rates, and the business value of shipped work. The report does not say that the 27 workplaces have adopted a common measurement framework.
Some respondents may face a near-term change as subsidized subscriptions end, while firms using API plans weigh spending caps, model choices, and open-model alternatives. Those changes could alter costs independently of how much code engineers produce. The next useful evidence would be comparable, time-stamped cost and quality measures—not PR volume alone—and the report leaves open whether companies will publish such results.
Key Questions
What did the report find about AI spending per engineer?
Spending varied widely across the 27 workplaces contacted. Examples ranged from a $150 monthly cap per developer at a large publicly traded company to a reported $3,000 per developer per month at a two-person .NET shop. The sample is not presented as representative of the whole industry.
Did AI tools cause the increase in pull requests?
The data show higher PR counts in the CodeRabbit and Gitar material, but they do not establish that AI tools caused the changes. The report does not isolate AI use from other factors that can affect output.
Does the report calculate AI’s return on investment?
No. It says few companies can quantify ROI so far. The material describes spending and activity, but does not calculate whether financial or productivity gains exceeded costs.
Why might more pull requests not mean better results?
PR volume measures how often changes are submitted, not their quality, review effort, defect rates, or business value. The report does not measure those factors, so higher volume alone cannot show that a team is getting a better return.
What could change companies’ AI bills next?
The report says some companies’ subsidized subscriptions are ending and that some are changing models or using open models. Those shifts could affect costs, but the source does not establish how bills will change across the industry.
Source: rss
