🔍 Read the full analysis: How To Find The Real Return On AI Coding on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
A report based on interviews at 27 workplaces finds AI coding budgets range from about $150 to $3,000 per developer per month, while measured output varies widely. The evidence points to gains in tasks such as issue triage and backlog cleanup, but does not establish a single reliable return across teams; companies need to measure delivery outcomes, rework and costs.
Engineering teams are spending about $150 to $3,000 per developer per month on AI coding tools, but many cannot yet say what that spending has delivered, according to an October 23 report by Thorsten Meyer that reviews data from 27 workplaces. The findings arrive as companies set 2027 budgets and face usage-based pricing, while reported productivity results range from flat output to large increases.
The report draws on interviews and data assembled by Gergely Orosz of The Pragmatic Engineer, who spoke with engineering leaders and senior engineers at 27 workplaces and included aggregate figures from code-review vendors CodeRabbit and Gitar. The freely accessible portion of Orosz’s report is the basis for Meyer’s account. Spending levels did not appear to track company size: a publicly traded company with 1,500 engineers capped individual spending at $150 a month, while a two-person European .NET agency reported spending about $3,000 per developer.
Productivity measures also diverged. CodeRabbit reported that developers on its platform created 113% more pull requests per week than a year earlier. Microsoft’s study of tens of thousands of engineers during an early-2026 rollout found adopters merged about 24% more pull requests than they otherwise would have over four months. Those figures measure different groups and outcomes; neither establishes how much additional customer or business value was delivered. At some interviewed companies, pull-request counts were flat, while one reported that pull requests had become three times larger.
Controlled research offers a more qualified picture. METR’s early-2025 study found experienced open-source developers took 19% longer with AI on the tasks tested. A February follow-up estimated an 18% speedup for returning developers and 4% for new developers, but the confidence intervals included zero. METR said the results were unreliable in part because participants held back some tasks they preferred to do without AI. It also reported that participants in a survey of 349 technical workers estimated a 1.4-to-2-times change in the value of their work, while warning that its 2025 participants had overestimated AI’s effect by 40 percentage points.
Infographic 01 · Where the money goes
From a $150 cap to $3,000 a month per developer.
Anecdotal data from 27 workplaces, mostly anonymized. It shows how wide the range is, not an industry average.
Spend per developer per month
Examples from the free part of the article · scale to $4,000
actual or typical spendbudget cap
What changed in 2026
Why the 2027 budget needs a range
- Peak in June and July: spend has roughly halved since for companies on API plans (Opus 5.5 and open models).
- Caps are common at medium and high spenders, and most people say they are easy to get around.
- Subsidies are ending: some companies must optimize or see the bill rise 5–10×.
- GitHub Copilot bills by token since 1 June 2026, at each model’s API rate, with no fallback to a cheaper model.
- Company size does not predict spend.
ROI is “a question that few companies seem able to answer – at least for now.”Gergely Orosz, The Pragmatic Engineer, 20 Oct 2026
Sources: The Pragmatic Engineer, “What’s the ROI of software engineers’ spending on AI?” (20 Oct 2026, free part); GitHub (27 Apr 2026).
Powered by Thorsten Meyer AI
Budget Decisions Need Delivery Measures
The immediate issue for companies is not whether AI tools can produce code, but whether the spending translates into accepted, useful changes without shifting costs into review, rework or security. That distinction matters as discounts expire and billing increasingly reflects model usage. The report says some companies expect discounted subscriptions to end soon; without cost controls, their bills could rise five- to tenfold, according to those companies.
Meyer points to cost per accepted change, rework, change-failure rates and the number of stories closed compared with new stories arriving as more useful measures than tokens consumed or code generated. These are suggested measures, not results from a standardized industry study. They help separate faster drafting from successful delivery and make it possible to compare tool costs with the work a team actually completes.
The gap is visible in the examples. One SaaS company saw little change in merged-pull-request velocity, yet a staff engineer described customer-issue triage shrinking from as much as a day to minutes of agent work and up to an hour of checking. A multinational team reportedly closed three times as many user stories per month. The results suggest that productivity gains may depend on the task and the team’s software environment, rather than following automatically from adoption.
As an affiliate, we earn on qualifying purchases.
Usage Pricing Meets Uneven Results
The budget question has sharpened as vendors change how they charge. According to Meyer’s account, GitHub moved all Copilot plans to usage-based billing on June 1, charging for tokens at each model’s API rate and removing the fallback to a cheaper model. GitHub said the change reflected higher compute and inference demands from agentic use. Orosz’s reporting also found API-plan spending peaked in June and July and had roughly halved since for some companies, which he attributed to Opus 5.5 and a shift to open models.
The reported use cases with clearer gains include issue triage, code archaeology, prototypes and backlog cleanup, particularly on simple, uniform technology stacks. Ix Technologies, which builds in C# and Blazor, reported shipping about 60 pull requests per developer per week while working through a long-standing backlog. Cofounder Rogier van der Hee said the two-person agency felt it could do the work of five to seven people compared with before AI. That is his account of the agency’s experience, not an independently verified productivity estimate.
Security remains another cost to track. Meyer cites Veracode testing in which 44% of AI coding tasks produced a known vulnerability. The figure is tied to that test and does not show that 44% of AI-generated code in all workplaces is vulnerable. It does underline why output counts alone cannot capture the review and security work needed before software is shipped.
““A question that few companies seem able to answer – at least for now.””
— Gergely Orosz, as quoted in Thorsten Meyer’s report
As an affiliate, we earn on qualifying purchases.
How Much Output Becomes Value
The available figures do not establish a common, independently verified return across companies. Studies and vendor data cover different users, tasks, time periods and definitions of productivity. A higher pull-request count may reflect more valuable work, larger changes, or additional review burden; the source material does not provide a consistent way to distinguish these outcomes across the 27 workplaces.
METR’s estimates also carry substantial uncertainty, and its researchers flagged limitations in how participants selected tasks. The Veracode result is specific to its test, while company anecdotes cannot show how typical those gains are. It is also unclear how often reported savings in drafting time persist after review, maintenance and security checks. Meyer’s report does not supply a standardized cost-benefit analysis or a complete dataset covering all participating workplaces.
AI-powered developer productivity tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Companies Set Their 2027 Baselines
Companies setting 2027 budgets will need to decide whether to renew discounted subscriptions, impose spending caps or shift workloads among models. The report says caps are common and can be bypassed, making policy alone an incomplete way to control costs. Teams will need to compare spending with delivery measures over a defined period and account for review, rework and failures alongside output.
The next useful evidence would show results on comparable tasks and include both costs and quality: accepted changes, time to delivery, rework, security findings and the rate at which teams close work relative to incoming demand. Until companies publish or collect that information consistently, the return on AI coding will remain a local measurement question rather than a settled industry figure.
software development analytics tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
How much are companies spending on AI coding tools?
The workplaces described in the report ranged from a $150 monthly cap per developer to about $3,000 per developer per month. These are examples from the reported sample, not an industry-wide average.
Does the report show that AI coding increases productivity?
Some measures show increases, including Microsoft’s reported 24% rise in merged pull requests among adopters over four months. Other companies reported flat pull-request counts, and METR’s research estimates were uncertain. More output does not by itself establish more business value.
Where does AI coding appear most useful?
The cases in the report point to triage, code archaeology, prototyping and backlog work, especially in simpler, more uniform software environments. These are reported experiences, not proof that the same gains will apply to every team.
What should engineering teams measure?
Meyer recommends tracking cost per accepted change, rework, change-failure rates and stories closed compared with stories coming in. Teams should also account for review and security work rather than treating generated code or token use as the return.
What remains unknown about AI coding’s return?
There is no consistent, cross-company measure showing how much reported output turns into delivered value after review, maintenance and security checks. The cited studies use different methods, and several findings carry significant limitations.
Source: ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
