What happened
In an October 20, 2026 article, Gergely Orosz of The Pragmatic Engineer examined what software organizations spend on AI tools and whether the spend has changed delivery output. The article draws on conversations with engineering leaders and senior-plus engineers at 27 workplaces, most anonymized at their request. It also includes aggregate data supplied by CodeRabbit and Gitar. This is a reported survey, not a controlled study, and the author frames it as an attempt to answer a practical budgeting question: what other companies spend per engineer, and what happened to pull-request throughput.
The spending range is wide. The article says company size and AI spend were unrelated in this data set. At one end, a two-person European .NET startup reported spending $3,000 per developer each month, mainly on Claude Opus and GitHub Copilot. Its cofounder said the team now makes about 60 PRs per developer per week, versus one or two per day previously, and attributed a 5-to-10-times increase in PR frequency to its use of AI tools. At the other end, the supplied summary identifies UK fintech Lendscape at $65 per developer per month. The article also describes a publicly traded company with 1,500 engineers that caps AI spending at $150 per developer monthly.
Several reported patterns complicate a simple “more spend, more output” interpretation. Medium and higher spenders commonly use caps, although respondents said the limits are often easy to avoid. Some companies have been using subsidized subscriptions, and several told the publication those subsidies are ending; they expect either to optimize usage or face bills five to ten times higher. The article says spending peaked during June and July at nearly all companies in its sample and roughly halved for API-plan users afterward. Respondents attributed much of that reduction to Opus 5.5 and moves toward open models. Claude and Opus 5.5 were the most frequently cited destinations for spend, while Cursor, Codex, and GitHub Copilot were also mentioned.
The largest total-spend example is a quantitative trading firm with about 5,000 staff and roughly 2,000 developers. It reportedly went from $500,000 in monthly AI spending in January to $8 million in September, described as a 16-fold increase. Average monthly spending was $1,600 per developer, up from around $80. An engineer said traders, rather than developers, were the biggest users and that the organization was learning to rate-limit expensive accidental queries. Its per-employee AI limit was reported as $25,000. For that firm, the article says ROI is evaluated at the trading-desk level through profit and loss, with AI counted as part of the cost base.
Output data offers another signal, but not proof of causation. CodeRabbit data showed a 113% increase in weekly PRs over a year, a 127% rise in monthly PRs, and a 128% rise when following the same 20,000 developers for a year. Gitar, described as Sonar’s AI code-verification agent, sees “AI-native” teams shipping more than 60 PRs per developer per week. Orosz explicitly notes that culture may affect those levels: teams acclimate to the pace expected around them. The central conclusion is consequently narrow: AI spending has increased, PR output has risen in some available data, but few companies can yet say what ROI actually is.
What the archive adds
This is a new fact on a known operating problem. Our earlier note, Cut your AI spend with AI Gateway's Auto Router, covered automatic request-level model routing as a cost-control mechanism. The new survey adds reported budget ranges, changing costs, and the harder question that routing alone cannot answer: whether lower spend produces accepted engineering outcomes.
Trend signal
The survey supports Agent economics moves to completed work. Its evidence does not establish a universal cost per completed task, but it makes the need for one clearer: monthly bills, model choice, and PR volume are all partial measures. A cheaper model can reduce a bill without preserving quality; a larger number of PRs can increase review demand without increasing delivered value.
Why now
The article places the reporting in 2027 budgeting season, after a summer spending peak. That timing matters because several respondents expect subsidized subscriptions to end, while API users have already seen costs fall after changing models or moving toward open alternatives. Procurement pressure is arriving before most teams have settled their measurement model.
Connections to other events
The archive’s AI Gateway note described automated, request-level routing among models. This survey supplies the organizational context for that infrastructure decision: routing may lower model cost, but it cannot on its own demonstrate ROI. The reported spending caps and rate limits also show that governance is becoming part of everyday AI-tool operations, not merely a finance exercise.
Why it matters
A team can take the article as a warning against two easy dashboards. The first is a vendor or API invoice, which measures consumption but not accepted work. The second is PR volume, which is readily available and may be rising, but can reflect smaller changes, different review conventions, or work shifted into cleanup. Neither tells a leader whether AI shortened cycle time, reduced defects, accelerated a backlog, or merely moved effort into specification and verification.
The survey also makes cost architecture a product decision. Model defaults, subscriptions, caps, routing, and rate limits now affect what builders can attempt and how reliably they can repeat it. The useful question is not whether a particular developer spends $150 or $3,000 monthly. It is whether an AI-assisted workflow completes a defined class of work at an acceptable total cost, including human review and rework.
What happens next
- More organizations will replace broad AI allowances with differentiated limits or routing policies. (6 months, 70%). Watch for: documented caps by role, model, or workload.
- Teams will add outcome measures alongside tool spend and PR counts. (12 months, 65%). Watch for: reporting on accepted tasks, cycle time, defects, or review effort.
Limits and uncertainty
The source is based on 27 reported company cases, many anonymized, plus aggregate platform data. It does not provide a representative sample, audited financial records, a common definition of “AI spend,” or controlled before-and-after comparisons. The PR figures cannot establish that AI caused the increase, and PRs are not equivalent to customer value, quality, or production impact. Individual company claims, including team-size equivalents and expected savings from model versions, are respondents’ accounts. The article also does not disclose enough detail to compare task mix, codebase age, review policies, or subscription terms across all organizations.
Skill to build now
Create a small weekly ledger for one recurring engineering workflow. Record the task type, AI-tool and model cost, time to an accepted merge, reviewer time, rework or rollback, and whether the task reached production. Keep PR count as a descriptive signal, not the result. After several comparable tasks, calculate a total cost per accepted outcome and compare it with the team’s prior baseline. This gives procurement, engineering, and finance one shared object for discussing limits or model routing.
Reader takeaway
AI coding expenditure is becoming material enough to manage, but the survey shows that spending and PR volume are not a return-on-investment model. Measure a bounded, accepted engineering outcome with its review and rework costs. That is the basis for deciding whether to raise a cap, change a model, or keep an AI workflow at all.
Open question
Which outcome metric can teams standardize across different codebases without rewarding superficial PR fragmentation or hiding additional verification work?