TL;DR
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
The Pragmatic Engineer gathered spending and pull-request data from 27 workplaces, alongside aggregate figures from CodeRabbit and Gitar. The figures suggest developers are producing more pull requests, but the report says few companies can yet answer whether AI tools’ productivity gains justify their costs.
The Pragmatic Engineer has published a survey of AI-tool spending at 27 workplaces, finding that companies pay sharply different amounts per engineer while many still cannot say whether the tools deliver a measurable return. Aggregate data shared by CodeRabbit shows weekly pull requests among developers using its platform rose 113% on average over a year, but the report cautions that output counts alone do not establish that AI caused the increase or that the spending paid off.
The survey combines interviews with engineering leaders and senior engineers at 27 workplaces with aggregate metrics from AI code-review and verification providers CodeRabbit and Gitar. The author anonymized most participating companies at their request. The report presents spending figures and changes in pull-request (PR) throughput as a snapshot of how organizations are evaluating AI coding tools, not as a controlled comparison.
CodeRabbit’s anonymized platform data showed a 113% increase in weekly PRs on average and a 127% rise in monthly PRs. Tracking the same group of 20,000 developers, the company reported a 128% increase over a year. Gitar, an AI code-verification agent, reported that some teams classified as “AI-native” ship more than 60 PRs per developer per week. These are vendor-supplied platform observations; the report does not establish that AI use alone produced the changes.
Spending varied widely, with no apparent relationship between company size and per-engineer outlay in the sample. Examples ranged from a two-person .NET business spending $3,000 per developer per month to a publicly traded company with 1,500 engineers that capped monthly AI spending at $150 per developer. The report says spending limits were common among medium and higher spenders, while some companies relied on subsidized subscriptions whose costs may rise when subsidies end.
Why Higher PR Counts Do Not Prove ROI
The findings show why companies are tracking PR throughput: it is a readily available measure that can be compared over time. If teams can submit and review more work without sacrificing quality, that could matter to delivery speed and engineering capacity. But a higher count does not, by itself, show that more valuable software was shipped, that work took less time, or that the additional output came from AI rather than other changes.
For employers, the gap between rising bills and uncertain returns makes measurement a practical budget issue. The source reports that some API-based costs peaked in June and July and later roughly halved at companies that changed models or moved to open models. Elsewhere, some subsidized subscriptions are nearing their end, potentially increasing costs by five to 10 times, according to the report. The range means a single spending benchmark is unlikely to fit every organization.
The report also describes a trading firm whose total AI spend reached $8 million a month in September, up from $500,000 in January. A current engineer told the author that traders were among the biggest users and that the company was learning to limit expensive queries. In such a setting, a firm may judge tools against trading outcomes rather than PR counts, underscoring that ROI depends on how and where the tools are used.
As an affiliate, we earn on qualifying purchases.
How Companies Are Tracking AI Use
The author began investigating after an engineer asked what other companies spend per developer and how PR throughput changed before and after AI spending. The report says spending on AI agents was growing among the technology companies the author spoke with, prompting organizations to seek more realistic estimates of their bills. It treats PR counts as a convenient comparison point, while acknowledging that teams have different norms for how much code they ship.
Among the 27 workplaces, the report describes large differences in budgets and controls. Some organizations set per-person caps and require additional spending to be justified; others offered generous allowances. A cited trading firm set a $25,000 per-employee limit, while the engineer interviewed said its recent profitability helped make AI spending affordable. The firm was also building internal computing capacity and models, according to the report.
Model and product choices also shaped bills. The source says Claude and Opus 5.5 were frequently named as major spending destinations in this limited sample, while Cursor, Codex and GitHub Copilot also appeared. It attributes the reported decline in some API costs partly to model changes and adoption of open models. These examples describe respondents’ experiences, rather than a broad market-wide accounting.
““Lots of traders are using Claude. Engineering is learning the importance of rate limiting for traders who accidentally run expensive queries without realizing it.””
— A current engineer at the trading firm, quoted by The Pragmatic Engineer
developer pull request management tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What the Survey Cannot Establish
The report does not provide a controlled test showing that AI tools caused the reported growth in PR counts. The platform figures cover users of particular products, and the source does not establish that those users represent software engineers generally. Nor does it show whether higher PR volume translated into better software, fewer defects, faster delivery or lower labor costs.
The 27-company sample is not presented as representative of the industry, and most participating companies were anonymized. The available material also does not provide a common accounting method for subscription charges, API usage, internal computing costs and employee time. As a result, comparing monthly spend across companies may omit different cost components.
The report says few companies seem able to answer the ROI question, at least for now. It does not give a single industry-wide return figure, nor enough detail to determine whether the biggest spenders gained more than lower-spending organizations. The timing and durability of reported spending changes also remain uncertain.
AI-powered coding verification tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Budgets Meet Better Outcome Measures
Companies facing the end of subsidized subscriptions will need to decide whether to absorb higher bills, restrict usage or change tools, according to the report. Organizations using API plans may also continue adjusting model choices and spending limits as costs shift. The trading firm cited in the report is building internal GPU and CPU capacity and developing its own models, though the source gives no timeline or independent assessment of those projects.
The next useful evidence would connect spending with outcomes beyond PR volume: review and rework rates, defects, delivery time and the value of work completed. Until companies publish comparable measurements, the report offers a picture of what organizations are spending and how platform activity has changed, but not a settled answer on whether AI coding tools pay for themselves.
software engineering productivity tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did the report find about AI spending per engineer?
Spending varied widely across the 27 workplaces surveyed. Examples ranged from $150 per engineer per month at one large company to $3,000 at a two-person .NET business; the report found no apparent link between company size and per-engineer spending in its sample.
Did pull requests increase after developers adopted AI tools?
CodeRabbit reported a 113% average increase in weekly PRs among developers using its platform, and a 128% rise when tracking the same 20,000 developers over a year. These observations do not prove AI caused the increase.
Does the report show that AI tools have a positive ROI?
No. The report says few companies can yet answer the ROI question. Higher PR counts are a measure of activity, but do not establish improved software quality, lower costs or greater business value.
Why might AI spending change?
The report cites model choices, movement to open models, usage limits and subscription subsidies as factors. Some companies said API costs had fallen from summer peaks, while some subsidized subscriptions were nearing their end.
Source: rss
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
