Key Takeaways #
- The largest matched study, of more than 100,000 developers, found a 180 percent rise in commits produced only 30 percent more releases and no measured increase in product use.
- Across 27 workplaces surveyed by The Pragmatic Engineer, AI coding spend ran from a $150 monthly cap to $3,000 per engineer, and bore no relation to company size.
- Pull request volume more than doubled in CodeRabbit's data, but several companies that kept their review standards saw no change in PR frequency.
- At $1,000 a month, a 20 percent release uplift pays back for most engineers. The larger risk is price: subsidised plans are ending, and Gartner expects AI coding costs to exceed developer salaries by 2028.
- Measure releases, lead time, review time and change failure rate. Pull request counts and self-reported time savings both overstate the gain.
Sources, checked 21 October 2026: Gergely Orosz, What's the ROI of software engineers' spending on AI?, The Pragmatic Engineer (20 October 2026; free section only); Writing Code vs. Shipping Code, NBER working paper 35275; Faros AI, The AI Productivity Paradox (July 2025) and AI Acceleration Whiplash; Google Cloud, 2025 DORA report; METR, early-2025 developer productivity trial; GDS, AI coding assistant trial findings; Gartner's 24 June 2026 forecast, as reported by IT Brief UK; Jensen Huang's token-spend remarks, as reported by R&D World (24 March 2026); Adam Mosseri's remarks on token caps, as reported by TechCrunch (14 July 2026); and our own systematic review of 126 AI coding productivity studies.
When developers in a study of more than 100,000 GitHub accounts moved from autocomplete to autonomous coding agents, their commits rose 180 percent. Their releases rose 30 percent. Across four application marketplaces, total usage of what they shipped did not rise at all.
That gap now has a price attached. Gergely Orosz's survey of 27 workplaces for The Pragmatic Engineer, published on 20 October, found companies spending up to $3,000 per engineer a month on AI coding tools. One quantitative trading firm went from $500,000 a month in January to $8 million in September. Most of the companies he spoke to judge the return by pull request counts, and finance teams are starting to ask what the extra spend buys.
The survey is the best current picture of what companies pay. Set against the wider research on what AI coding delivers, it shows where the return is real and where it is an artefact of measurement.
How Much Do Companies Spend on AI Coding per Engineer? #
Spend per engineer varies by a factor of 20. Orosz draws five spending trends from his interviews:
- Spend does not track company size. The highest monthly spend per engineer in the sample, $3,000, is at a two-person .NET startup. A public company with 1,500 engineers caps spend at $150.
- Spend peaked in the summer. June and July were the peak months almost everywhere. Since then, costs on API plans have roughly halved, mostly because teams moved to Claude Opus 5.5 or to open-weight models.
- Caps are normal above modest spend levels. Extra spend usually needs a justification, although most interviewees said the caps are easy to get around.
- Subsidies are ending. Flat-rate subscriptions hide the real cost of usage. Several companies whose subsidised plans are expiring expect either to cut usage or to see bills rise five to 10 times.
- Claude takes most of the spend. Anthropic's models, Opus 5.5 in particular, dominated answers about where the money goes, with Cursor, Codex and GitHub Copilot also in use. Orosz notes the sample is small.
His interviewees, mostly anonymised, reported the following spend and output figures.
| Organisation | Spend per engineer a month | Reported change in output |
|---|---|---|
| Two-person .NET startup, EU | $3,000 | PR frequency up five to 10 times; around 60 PRs per developer a week |
| Public company, 25,000+ staff | $1,000 average | PRs up from about four to eight per developer a week |
| Multinational, 10,000+ staff | $1,000 average, $4,000 hard cap | Closed user stories up three times; one team's PRs from 1.5 to four a week |
| Public SaaS company, 600 engineers | $1,200 budget; fewer than 5 percent reach it | More PRs, little change in merge velocity |
| Series D hospitality company, 200 engineers | $1,500 default cap | PR frequency roughly unchanged |
| Public company, 5,000 staff | $1,000 soft cap; $1,500 to $2,500 actual | PR frequency flat; PRs three times larger |
| AI SaaS company, 100 staff | $2,000 effective hard cap | Velocity largely unchanged outside greenfield work |
Aggregate tool data looks stronger. CodeRabbit told Orosz that weekly pull requests among its users rose 113 percent year on year, and 128 percent across a fixed cohort of 20,000 developers. Gitar, Sonar's code verification agent, reports "AI-native" teams at more than 60 PRs per developer a week.
The pattern inside the table matters more than the headline. The companies reporting flat PR frequency are the ones that kept their previous standards for PR size and quality. The hospitality company built an agent to flag oversized or verbose changes.
Why Do Pull Requests Overstate AI Coding ROI? #
Orosz chose pull requests because almost every company records them and few have yet had reason to game them. That makes them a fair before-and-after comparison of activity. The wider evidence shows activity and delivered value diverge sharply.
The clearest test is the study behind the opening figures, Writing Code vs. Shipping Code, an NBER working paper that used matched event studies. Autocomplete, interactive agents and autonomous agents were associated with cumulative commit increases of 40, 140 and 180 percent. The 180 percent rise in commits became 50 percent more projects and 30 percent more releases. Across four application marketplaces, it produced no measured increase in total usage.
Commercial telemetry shows where the extra code goes. Faros AI's 2025 analysis of more than 10,000 developers on 1,255 teams found that high-adoption teams merged 98 percent more pull requests and completed 21 percent more tasks. Their review time rose 91 percent, average PR size rose 154 percent, and no organisation-level improvement in DORA delivery metrics appeared. Faros's later two-year study of 22,000 developers reported 861 percent more code churn and 242.7 percent more incidents per pull request. Both are vendor datasets, but they agree with Google's 2025 DORA report: in a survey of nearly 5,000 technology professionals, AI adoption improved delivery throughput while continuing to reduce delivery stability.
Self-reported gains are no more reliable. In METR's randomised trial, 16 experienced open-source developers took 19 percent longer on tasks where AI was allowed, yet afterwards believed it had made them 20 percent faster. The study used early-2025 tools and a small sample, so the slowdown should not be generalised. The perception gap should. The UK government's own AI coding assistant trial, which issued 1,900 licences across more than 50 public sector organisations, reported 56 minutes saved per developer per day. That figure was self-reported.
More code is not more software. The bottleneck moves to review, integration and operations, and pull request counts stop measuring anything useful at the point where the cost starts to fall on other people.
The Arithmetic Still Favours Moderate Spend #
None of this means the spend is wasted. Our systematic review of 126 studies found controlled coding-stage gains of roughly 20 to 30 percent, and concluded that 10 to 30 percent higher release output, centred on 20 percent, is the defensible planning range. Orosz groups his companies into high, medium and low spenders. The arithmetic differs for each.
| Spend tier | Per engineer a month | Per year | Break-even loaded cost at a 20 percent uplift | Break-even at five times the price |
|---|---|---|---|---|
| High | $1,000 to $3,000 | $12,000 to $36,000 | $60,000 to $180,000 | $300,000 to $900,000 |
| Medium | $300 to $500 | $3,600 to $6,000 | $18,000 to $30,000 | $90,000 to $150,000 |
| Low | Under $300 | Under $3,600 | Under $18,000 | Under $90,000 |
High spenders: $1,000 or more
Most of Orosz's detailed examples sit here, and this is where the case is least certain. At $1,000 a month, break-even needs a fully loaded cost of $60,000, which covers most professional developers in the UK and US. At $3,000 it needs $180,000 and an above-median gain. A small team doing work that once took several times as many people plausibly clears it. A large engineering organisation with flat release output does not. This tier also has no headroom: at five times current prices, no ordinary salary covers the cost.
Medium spenders: $300 to $500
Orosz profiles five companies in this range. At these levels, a 20 percent uplift pays back for any engineer costing more than $18,000 to $30,000 a year, so the return is close to certain if the uplift is real. A fivefold price rise moves the threshold to $90,000 to $150,000. That still covers many senior engineers, but not every team.
Low spenders: under $300
Many companies in this tier rely on subsidised flat-rate subscriptions, which makes the cost look trivial. The exposure is the subsidy, not the spend. If bills rise five to 10 times when subsidies end, as several of Orosz's interviewees expect, a $200 seat becomes $1,000 to $2,000 a month. That moves a company into the high tier without any change in how its engineers work.
Where prices are heading
The weaker part of the calculation is the price. Spend fell this summer because model costs fell, but subsidised plans are ending.
Gartner predicted on 24 June that AI coding costs will exceed the average developer's salary by 2028, as usage-based pricing replaces seats and agentic workflows consume more tokens. "Token discipline will not emerge through developer choice alone," said Nitish Tyagi, a senior principal analyst at the firm. A return that holds at $1,000 a month may not hold at five times that.
Big Tech already shows both ends of the debate. On the All-In Podcast during NVIDIA's GTC conference in March, Jensen Huang said an engineer paid $500,000 a year should be using at least $250,000 of tokens. Four months later, Meta's Adam Mosseri told Lenny's Podcast that a strong engineer's token spend might soon match their cost of employment, and that "you're going to probably need to put in some caps". Meta has no caps yet, but has already shut down an internal leaderboard of token spend. If the companies with the most room to spend are preparing per-engineer budgets, smaller buyers should plan for the same.
Five Steps to Measure the Return Before Finance Asks #
The trading firm in Orosz's survey judges its AI bill by one number: each desk's profit. That works because desk profit is an outcome, not an activity. Most engineering organisations cannot tie code to revenue that directly, but they can measure much closer to delivery than PR counts. In order:
- Set a delivery baseline. Record releases, lead time for changes and change failure rate before rollout, or reconstruct them from the past year. These are the DORA metrics most teams already collect, and they are where the NBER and Faros results part company with PR counts.
- Hold the quality bar constant. Keep existing limits on PR size and scope, as the hospitality company in Orosz's survey did. A flat PR count under unchanged standards tells you more than a doubled count under relaxed ones.
- Report review time next to output. If merged PRs and review time rise together, the gain has moved to reviewers rather than reached customers. Do not substitute self-reported time savings: METR's results show developers overestimate them.
- Cap spend with cheap escalation. Use soft limits, a fast approval route and monitoring of high-token workflows, as Gartner recommends, with smaller models for simple, frequent tasks. Hard caps that engineers route around measure nothing.
- Test the business case at higher prices. Rerun it at three and five times current spend, because subsidised plans are ending. OurAI coding cost calculator andtool price tracker cover the inputs.
We track the evidence on AI coding productivity and cost as it is published. Get new posts by email.