AI Coding Economics: Where The Money Goes And What Comes Back

🔍 Read the full analysis: AI Coding Economics: Where The Money Goes And What Comes Back on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get hardware and tech essentials delivered free with Prime

• Fast, free delivery on millions of items
• Prime Video, Amazon Music and more included
• Member-only deals all year

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

Engineering teams are reporting AI coding costs from $150 to about $3,000 per developer per month, while evidence of productivity gains varies widely, according to an October 20 report by Gergely Orosz in The Pragmatic Engineer. The findings, based on conversations at 27 workplaces and data from two code-review vendors, arrive as companies plan 2027 budgets and face the question of whether higher spending produces more valuable software.

Orosz’s report combines interviews with engineering leaders and senior engineers at 27 workplaces with aggregate figures from CodeRabbit and Gitar. The published article says company size did not reliably predict spending. Examples range from a publicly traded company with 1,500 engineers capping monthly use at $150 per developer to a two-person European .NET agency that spends about $3,000 per developer. Some organizations set caps between $1,000 and $4,000, with typical use around $1,000.

Reported output measures do not tell a consistent story. CodeRabbit says weekly pull requests on its platform rose 113% year over year; that figure reflects teams using its AI code-review product. A Microsoft study of tens of thousands of engineers during an early-2026 rollout found adopters merged about 24% more pull requests over four months than they otherwise would have. Microsoft’s researchers cautioned that a merged pull request does not necessarily represent delivered value. At companies in Orosz’s survey, pull-request activity ranged from flat to several times higher, and one company saw larger pull requests without more frequent submissions.

The report also describes a quant trading firm whose monthly AI spending grew from $500,000 in January to $8 million in September. The source says its largest spenders were traders, and the firm evaluates costs against desk-level profit and loss. That kind of direct financial measure is unusual among the companies interviewed: Orosz writes that few could answer whether their AI spending was producing a return.

How AI Costs Meet Engineering Results

The gap between software activity and business value is central to the budget question. More pull requests or faster coding can be useful, but they do not by themselves show whether a company ships more customer value, reduces delays or spends less overall. Orosz’s report recommends tracking measures such as cost per accepted change, rework, change-failure rates and whether teams close more stories than enter their queues.

Some reported gains are tied to specific tasks rather than a broad increase in engineering output. At a SaaS company with about 600 engineers, a staff engineer said customer-issue triage that could take a day was reduced to minutes of agent work followed by up to an hour of human review. The same engineer described researching why code had been written a certain way as taking minutes rather than a day or two. These are interview accounts, not controlled measurements, but they show where teams say the tools save time.

Costs may also change as subsidies end and billing shifts toward usage. The source reports that several companies expected discounted subscriptions to expire, potentially increasing bills sharply if use stayed constant. That makes separating useful work from token consumption a practical budget issue, not just a productivity debate.

Why Spending Is Being Reassessed

Orosz reports that spending on API plans peaked in June and July and had roughly halved by the time of his report among some companies, which he attributes to the release of Opus 5.5 and moves to open models. This is a reported pattern among the companies discussed, not a comprehensive market-wide spending index. Some firms use monthly caps, though interviewees said those limits could be bypassed.

Billing policies are changing as AI tools handle more agentic tasks. GitHub said it 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. The company explained the change by saying, “Agentic usage is becoming the default” and brings higher computing and inference demands. The shift can make costs more directly reflect usage, while leaving employers to judge whether that usage is worthwhile.

Productivity research cited in the report is mixed. METR’s early-2025 study found experienced open-source developers took 19% longer with AI on selected tasks. A February follow-up estimated an 18% speedup for returning developers and 4% for new developers, but both confidence intervals included zero. METR said the data was unreliable in part because 30% to 50% of developers held back tasks they did not want to do without AI. The results underline that findings from particular tasks or teams should not be treated as a universal productivity rate.

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

— Gergely Orosz, The Pragmatic Engineer

What the Productivity Figures Cannot Show

The available figures do not establish a single productivity gain that applies across engineering teams. Vendor data comes from users of those vendors’ tools; interview examples are self-reported; and pull-request counts do not capture code quality, maintenance costs or customer outcomes. METR’s estimates also carry uncertainty, with confidence intervals that included no effect in its February follow-up.

The report does not provide a full, comparable accounting of tool expenses, staff time spent reviewing AI-generated work, security fixes or downstream failures across all 27 workplaces. Nor does it establish that the firms with higher output measures achieved proportionally greater business value. Veracode’s latest test, as cited by the source, found that 44% of AI coding tasks produced a known vulnerability; the source does not provide enough detail here to treat that result as a rate for all AI-generated code in production.

It is also unclear how widely the spending reductions reported for API plans apply, how many companies will lose discounts, and what their bills will be after that happens. The survey offers examples rather than a representative estimate of spending across the software industry.

How Companies Can Test Their Returns

The immediate next step for companies is to set budgets against measures of completed work and its quality, rather than counting tokens or generated code alone. The report points to tracking accepted changes, rework, change failures and backlog movement, alongside the cost of the tools and human review. Those measures can help reveal whether savings in one task are offset elsewhere.

As 2027 plans are finalized, companies will also need to account for subscription discounts ending and usage-based billing. The next useful evidence will be comparable results across teams and time: what work was completed, what it cost, how much review or repair it required, and whether delivery improved. Until organizations publish or collect that information consistently, claims of broad returns will remain difficult to verify.

Key Questions

How much are companies spending on AI coding tools?

The workplaces in The Pragmatic Engineer’s report described spending or caps ranging from $150 to about $3,000 per developer per month. The examples are not a representative industry-wide average.

Do the reported pull-request increases prove AI improves productivity?

No. The figures show changes in pull-request activity or merges in specific settings. Microsoft’s researchers said a merged pull request is not the same as the value it delivers, and vendor figures reflect users of those platforms.

Where did teams report the clearest gains?

Interviewees described gains in customer-issue triage, code archaeology, prototyping and backlog work, particularly on simple, uniform technology stacks. These examples are reported experiences, not controlled results across all companies.

What remains uncertain about AI coding ROI?

Many companies lack comparable data on tool costs, review and rework, software quality, failures and business outcomes. The reported studies also differ in their methods and do not establish one productivity rate for all engineering teams.

Source: ThorstenMeyerAI.com