AI Coding Budgets: Where Costs Rise And Value Shows Up

🔍 Read the full analysis: AI Coding Budgets: Where Costs Rise And Value Shows Up on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get the little things that make your day 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 reported spending from $150 to about $3,000 per developer per month on AI coding tools, according to a report based on interviews and data from 27 workplaces. The findings arrive as companies set 2027 budgets, but the evidence on productivity is mixed and many employers cannot yet show how tool use translates into business value.

The October 20 report by Gergely Orosz of The Pragmatic Engineer combined interviews with engineering leaders and senior engineers at 27 workplaces with aggregate data from AI code-review vendors CodeRabbit and Gitar. The source article says the figures cited are from the report’s freely available portion. Spending varied sharply: one publicly traded company with 1,500 engineers capped monthly use at $150 per developer, while Ix Technologies, a two-person European .NET agency, spent about $3,000 per developer.

Several output measures point in different directions. 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 during an early-2026 rollout found adopters merged roughly 24% more pull requests over four months than they otherwise would have, according to the report. The Microsoft researchers cautioned that a merged pull request does not by itself measure the value delivered.

Other workplaces reported flat pull-request frequency, larger changes, or gains ranging from doubled output to five- or tenfold increases. In controlled research, METR’s February follow-up estimated an 18% speedup for returning developers and 4% for new developers, but both estimates had confidence intervals that included zero. METR also described the data as unreliable, citing task-selection effects among participants. These measures are not directly comparable: they cover different teams, periods, tools and definitions of productivity.

Why Budget Owners Need Outcome Measures

The spending spread matters because a tool budget does not show whether a company is getting more useful work. Some firms have imposed monthly caps, while the report says those caps are often easy to bypass. It also describes discounted subscriptions nearing expiration at several companies; those employers may need to change usage or face bills that, according to the source, could rise five- to tenfold.

The clearest reported gains came from particular tasks, not from a universal increase in software delivery. At a SaaS company with about 600 engineers, a staff engineer said customer-issue triage that could take a day sometimes became minutes of agent work plus up to an hour of human checking. The same engineer described code-history research taking minutes instead of a day or two, and a product prototype taking hours rather than one to two weeks. These are company and employee reports, not controlled estimates of overall return.

The practical implication is to track outcomes such as cost per accepted change, rework, change failure rates and whether teams close more user stories than enter the backlog. Pull-request counts and token consumption can help describe activity, but neither alone establishes quality, reliability or business impact.

How AI Coding Costs Are Changing

The report says API-plan spending peaked in June and July and had roughly halved by the time of its account among companies that changed models or moved to open models. Orosz attributed that decline to the release of Opus 5.5 and adoption of open models. The report also describes a quant trading firm with about 2,000 developers whose monthly AI spending rose from $500,000 in January to $8 million in September. At that firm, traders, rather than engineers, were the biggest spenders, and desk-level profit and loss provided a measure for evaluating the expense.

Billing policies are changing alongside usage. The source says GitHub moved every Copilot plan 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 explained the change by saying, “Agentic usage is becoming the default, and it brings significantly higher compute and inference demands.” That shift makes the connection between usage, costs and measurable results more pressing for budget owners.

Evidence from individual companies also suggests results depend on the work and the software environment. Ix Technologies, which uses a uniform C# and Blazor stack, says it is clearing a backlog that reaches back about a decade. A multinational team reported closing three times as many user stories per month and, for the first time, more stories than were arriving. Those cases show reported results at particular organizations, not a guaranteed effect across engineering teams.

““Agentic usage is becoming the default, and it brings significantly higher compute and inference demands.””

— GitHub

What the Productivity Numbers Cannot Show

The reported results do not establish a single productivity effect across software teams. Vendor platform figures, company interviews, internal studies and controlled experiments use different samples and measures. CodeRabbit’s pull-request data comes from teams already using its review product, while PR volume does not establish whether changes are useful, maintainable or safe.

METR’s estimates remain uncertain: the confidence intervals in its February follow-up included zero, and the organization flagged limitations in how participants selected tasks. The source also says a METR survey found a median self-reported 1.4x to 2x change in the value of work, while METR noted that participants in its 2025 research had overestimated AI’s effect by 40 percentage points. Those self-reports should not be treated as verified productivity gains.

The source excerpt does not provide a complete cost-and-outcome dataset for the 27 workplaces, nor does it quantify the security and review costs across those companies. It cites a separate Veracode test in which 44% of AI coding tasks produced a known vulnerability, but the excerpt gives no further test details or comparison baseline. It is therefore unclear how that result applies to the workplaces surveyed or how often vulnerabilities were caught before release.

What Companies Should Track in 2027

As companies set 2027 budgets, the next test is whether they can connect tool spending to accepted, reliable work. The source recommends measuring cost per accepted change, rework, change failure rate and the balance between stories completed and stories arriving. These measures would give budget owners more information than token use or pull-request totals alone, although the article does not report that companies have adopted a shared standard.

Employers facing expiring discounts or usage-based bills will also need to monitor how model choice and agent use affect monthly costs. The report describes spending shifts after changes in models and plans, but does not provide a forecast for industry-wide expenditure or say which approach will deliver the best return. Further evidence from consistent, quality-aware comparisons will be needed to establish whether reported gains persist across different teams and codebases.

Key Questions

How much are companies spending on AI coding tools?

The workplaces in the report ranged from a $150 monthly cap per developer at one large company to about $3,000 per developer at a small European agency. The report found no simple relationship between company size and spending.

Does the evidence show that AI coding makes developers more productive?

Some measures show higher output, including more pull requests among CodeRabbit users and more merged pull requests among Microsoft adopters. Other companies reported flat counts, and METR’s estimates had confidence intervals that included zero. The evidence does not establish one productivity gain that applies to all teams.

Where did teams report the clearest benefits?

Reported gains were clearest in customer-issue triage, understanding older code, prototyping and clearing backlogs on relatively uniform software stacks. The examples are accounts from particular teams and should not be read as guaranteed results.

What should companies measure to judge AI coding costs?

The source recommends tracking cost per accepted change, rework, change failure rates and whether completed user stories outnumber incoming ones. These measures aim to connect spending with delivered work rather than usage volume alone.

Are AI coding tools creating security risks?

The source cites a Veracode test in which 44% of AI coding tasks produced a known vulnerability. It does not provide enough detail in the excerpt to determine how that finding compares with non-AI coding or how broadly it applies to companies in the budget survey.

Source: ThorstenMeyerAI.com