Where AI Coding Delivers Returns And Where It Drains Budget

🔍 Read the full analysis: Where AI Coding Delivers Returns And Where It Drains Budget on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on office and shipping supplies

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

A report published October 20 by The Pragmatic Engineer finds that AI coding budgets range from about $150 to $3,000 per developer per month across 27 workplaces, while companies report widely different changes in software output. The findings matter as organizations set 2027 budgets and face usage-based billing, but many still cannot say how much business value their spending produces.

The report draws on conversations with engineering leaders and senior engineers at 27 workplaces, plus aggregate data from AI code-review vendors CodeRabbit and Gitar. Its author, Gergely Orosz, says spending varies substantially even among companies of different sizes. Some set monthly caps; the report describes a publicly traded company with 1,500 engineers capping spend at $150 per developer, while a two-person European software agency spends about $3,000 per developer.

Output measures do not point to one consistent result. CodeRabbit reported that developers on its platform created 113% more pull requests per week than a year earlier. A Microsoft study of tens of thousands of engineers found adopters merged about 24% more pull requests over four months than they otherwise would have, according to the report. Those figures come from different settings and are not direct measures of the same outcome.

Some workplaces reported little or no change in pull-request frequency, though individual changes became larger. The report also cites METR research with mixed results: an early-2025 study found experienced open-source developers took longer with AI, while a February follow-up estimated speedups for returning developers and new developers whose confidence intervals included zero. METR cautioned that the follow-up data was unreliable.

Where Teams See Practical Gains

The reported benefits are clearest in specific, bounded tasks, rather than a general promise of faster software delivery. Engineers cited using AI to triage customer issues, understand unfamiliar code and produce prototypes. At one company with about 600 engineers, a staff engineer said issue triage that could take a day sometimes became minutes of agent work followed by up to an hour of human checking. The same engineer said tracing why code had been written a certain way could shrink from a day or two to minutes.

Teams working on a simple, uniform technology stack also described gains. Ix Technologies, a small agency that works in C# and Blazor, reported shipping about 60 pull requests per developer per week and clearing old backlog work. Cofounder Rogier van der Hee said the two-person team felt able to do work that would previously have required five to seven people. These are company-reported experiences, not controlled comparisons or proof that similar gains will apply elsewhere.

Spending can rise without equivalent value if teams optimize for usage or output counts rather than useful, accepted changes. The source reports that some organizations see larger pull requests without more frequent ones, and that review and security capacity can lag behind code production. It cites Veracode testing in which 44% of AI coding tasks produced a known vulnerability. That finding underscores why output volume alone does not establish a net return.

For budget owners, the report points toward measures such as cost per accepted change, rework, change-failure rate and whether completed user stories outpace incoming work. These metrics connect tool costs to software that teams accept and maintain, rather than tokens consumed or code generated. The article does not establish that any single metric or measurement method works across all organizations.

Billing Changes Put Budgets Under Scrutiny

The timing reflects both upcoming budget decisions and changes in how AI tools are charged. Orosz reported that spending on API plans peaked in June and July and had roughly halved by the time of his October report at some companies. He attributed that change to the availability of Opus 5.5 and a shift toward open models. The source does not provide a uniform spending series for all 27 workplaces.

The report says several companies were approaching the end of discounted subscriptions and expected costs could rise five- to tenfold if they did not optimize. Those are company expectations reported by Orosz, not confirmed future price increases. It also notes that spending caps are common and, according to some accounts, easy to bypass.

GitHub changed every Copilot plan to usage-based billing on June 1, charging tokens at each model’s API rate and removing the fallback to a cheaper model, according to the source. GitHub said agentic use was becoming the default and required substantially more computing and inference. That change gives organizations a more direct reason to track usage alongside outcomes, but does not by itself show whether a team’s investment is worthwhile.

“A question that few companies seem able to answer – at least for now.”

— Gergely Orosz, The Pragmatic Engineer

Productivity and Security Still Vary

The report does not establish a single productivity effect that applies across software teams. Its cited findings use different populations, time periods and methods: vendor platform data, a Microsoft study, workplace interviews and controlled research. Some figures measure pull requests, while others concern tasks or developers’ reported experience. Those measures cannot be treated as interchangeable, and the report’s author presents his view that individual output is probably up by double digits as an interpretation rather than a settled result.

It is also unclear how much of any time saved becomes higher-quality, shipped software after review, rework and maintenance. The source does not give a comparable cost-per-accepted-change figure across the 27 workplaces or a full account of the study methods behind every company example. The Veracode vulnerability result is a finding from its test, not a measure of the share of AI-generated code that is vulnerable in all production environments.

Future spending is uncertain as well. Subscription discounts may end, usage-based charges can change bills, and teams may switch models. The report’s budget examples show a wide range, but do not establish what a typical organization should spend or what return it should expect.

Track Outcomes Before Renewals

Companies setting 2027 budgets will need to decide whether to renew subscriptions, set usage limits, change models or fund additional review and security work. The report does not identify a single next industry milestone or a forthcoming follow-up study. For now, the practical next step it describes is to compare AI costs with delivery outcomes and quality measures at the team level.

That comparison can include cost per accepted change, rework, change failures and the balance between stories completed and new work arriving. Companies will also need to distinguish results by task and codebase: gains reported in triage or uniform stacks may not carry over to complex legacy systems. Whether teams can show sustained improvements after discounts end—and after review and maintenance costs are counted—remains to be seen.

Key Questions

How much are companies spending on AI coding tools?

The report describes budgets ranging from about $150 to $3,000 per developer per month across 27 workplaces. It does not identify one typical amount that applies across companies.

Does AI coding increase software output?

The findings are mixed. CodeRabbit reported a 113% rise in weekly pull requests on its platform, and a Microsoft study found adopters merged about 24% more pull requests over four months. Other companies reported flat counts, and pull-request volume does not by itself measure software value.

Where did teams report the clearest benefits?

Examples in the report include customer-issue triage, code archaeology, prototyping and backlog work on simple, uniform technology stacks. These are reported company experiences, not guaranteed results for other teams.

What should companies measure to judge return?

The source recommends looking beyond tool usage or code volume to measures such as cost per accepted change, rework, change-failure rate and whether completed user stories exceed incoming work.

What remains uncertain about AI coding costs?

It is not yet clear how much reported time savings translate into durable business value after review, rework, security and maintenance are counted. Future bills may also change as discounts end, usage-based billing applies and organizations switch models.

Source: ThorstenMeyerAI.com