By Thorsten Meyer | October 23, 2026 | ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Engineering teams now spend anywhere from $150 to $3,000 per developer per month on AI coding tools. Budgets for 2027 are being set right now, and most companies still can’t say what the money bought. Here is what the newest data says about where AI coding pays, where money burns, and how to tell the difference.
Key takeaways
- Spend is all over the map. Across 27 workplaces surveyed by The Pragmatic Engineer, budgets run from a $150 cap to about $3,000 per developer a month. Company size does not predict where a company lands.
- Output is up, but nobody agrees by how much. Merged PRs rose about 24% at Microsoft, weekly PRs more than doubled on CodeRabbit’s platform, and PR counts stayed flat at several companies.
- It clearly pays for triage, code archaeology, prototypes and backlog cleanup on simple stacks, and least on complex legacy code.
- Money burns where tokens become a target, PRs get bigger instead of better, and review and security can’t keep up. In Veracode’s latest test, 44% of AI coding tasks produced a known vulnerability.
- Measure outcomes, not consumption: cost per accepted change, rework, change failure rate, and stories closed versus stories coming in.
As an affiliate, we earn on qualifying purchases.
Why the question is landing now
On October 20, Gergely Orosz of The Pragmatic Engineer published the most concrete look at AI coding budgets I have seen this year. He talked to engineering leaders and senior engineers at 27 workplaces and added aggregate data from two AI code-review vendors, CodeRabbit and Gitar. Part of the piece is paywalled; everything I cite comes from the freely readable part.
Spending peaked in June and July and has roughly halved since for companies on API plans, which Orosz attributes to Opus 5.5 and a move to open models. Spending caps are common and, by most accounts, easy to get around. And subsidies are ending: several companies said their discounted AI subscriptions run out soon, so they must optimize or watch the bill rise five- to tenfold.
Vendors are moving too. On June 1, GitHub switched every Copilot plan to usage-based billing, charging tokens at each model’s API rate and dropping the fallback to a cheaper model. GitHub’s reason: “Agentic usage is becoming the default, and it brings significantly higher compute and inference demands.” Whoever owns the budget now needs a number.
As an affiliate, we earn on qualifying purchases.
Where the money goes
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
A publicly traded company with 1,500 engineers caps developers at $150 a month. A two-person .NET agency in Europe, Ix Technologies, spends about $3,000 per developer. In between sit companies with caps of $1,000 to $4,000, where the typical developer spends around $1,000.
The extreme case is a quant trading firm with roughly 2,000 developers. Its AI spend rose from $500,000 a month in January to $8 million in September, and the biggest spenders are traders, not engineers. That firm does not agonize over ROI: each desk’s profit and loss is the measure, and AI is part of the cost base. Most companies have no such scoreboard. Orosz’s own summary calls ROI “a question that few companies seem able to answer – at least for now.”
As an affiliate, we earn on qualifying purchases.
What the output data says, and where it disagrees
Infographic 02 · Who measured what
Output is up. Nobody agrees by how much.
Pull requests are easy to count and easy to misread. Here is what each source measured, and how much weight it can carry.
Same question, very different answers
Negative time = faster. Read the evidence type before the number.
| Result | Who measured what | Evidence type |
|---|---|---|
| +113% | CodeRabbit Weekly PRs per developer on its AI review platform vs. a year ago (+128% for one 20,000-developer cohort) | vendor data |
| +24% | Microsoft Merged PRs of tens of thousands of engineers adopting Claude Code or Copilot CLI, early 2026, vs. expected without | observational study |
| flat to 5–10× | 27 companies PR frequency per developer; one company flat while PR size tripled | interviews |
| −18% time | METR Task time for returning developers, Aug 2025 onward (CI −38% to +9%; new developers −4%) | RCT, not significant |
| +19% time | METR Same design, early 2025, experienced open-source developers | RCT, 2025 tools |
| 1.4–2× | METR survey Self-reported change in the value of work, 349 technical workers | self-report |
| −17% score | Anthropic study Comprehension quiz after learning a new library with AI (52 developers, GPT-4o) | RCT, learning |
Why they disagree: PR counts reward smaller or more PRs, not value; vendors measure their own users; METR lost the tasks developers refused to do without AI, so its estimate is likely a lower bound; self-reports overstated AI’s effect by 40 points in METR’s 2025 study.
Sources: The Pragmatic Engineer, “What’s the ROI of software engineers’ spending on AI?” (20 Oct 2026, free part) (incl. CodeRabbit data); Murphy-Hill et al., Microsoft (arXiv, 1 Jul 2026); METR (24 Feb 2026); METR survey (11 May 2026); Shen & Tamkin, Anthropic (arXiv, 28 Jan 2026).
Powered by Thorsten Meyer AI
- CodeRabbit says developers on its platform create more than twice as many PRs per week as a year ago (+113%). That is vendor data from teams already using an AI review tool.
- Microsoft studied tens of thousands of its engineers during the early-2026 rollout of Claude Code and GitHub Copilot CLI. Adopters merged roughly 24% more PRs than they would have otherwise, over four months. The authors add that “a merged PR is not the same as the value it delivers.”
- The 27 companies range from flat to five- to tenfold. At one, PR frequency stayed flat while PRs became three times larger. At another, PRs doubled from about four to eight per developer per week.
Controlled experiments are messier. In early 2025, METR found experienced open-source developers took 19% longer with AI. Its February follow-up estimated an 18% speedup for returning developers and 4% for new ones, but both confidence intervals include zero, and METR calls the data unreliable because 30% to 50% of developers held back tasks they didn’t want to do without AI. Self-reports run higher: in a METR survey of 349 technical workers, the median reported a 1.4x to 2x change in the value of their work, though METR notes that its 2025 participants overestimated AI’s effect by 40 percentage points.
My reading: individual output is up, probably by double digits. How much becomes delivered value depends on the team, not the tool.
software development productivity tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Where AI coding clearly pays
The most revealing case is a SaaS company with about 600 engineers where merged-PR velocity barely moved. A staff engineer there still named clear wins: triaging customer issues went from sometimes a whole day to minutes of agent work plus at most an hour of human checking. Finding out why code was written a certain way, from Slack, design docs and PR history, takes minutes instead of a day or two. A complex prototype for product managers takes hours instead of one to two weeks.
The other clear win is backlog work on a simple, uniform stack. Ix Technologies builds everything in C# and Blazor and ships about 60 PRs per developer per week, clearing a backlog that goes back about ten years. Cofounder Rogier van der Hee says it feels like the two of them do the work of five to seven people before AI. At a multinational, one team now closes three times as many user stories per month, and for the first time more than come in.
DORA’s ROI report cites Stanford research with the same pattern: a 35% to 40% gain on simple, greenfield tasks, often 10% or less on complex legacy code. I have that figure only as cited by DORA and reported by InfoQ.
Where the money burns
Tokens as a target. Some companies now rank employees by token consumption. DORA calls it “tokenmaxxing” and warns of “productivity theater”, with developers running agents on nonsense projects to stay above the company average. It also cites Adnan Masood’s point that the highest budgets often buy only a marginal throughput increase at ten times the cost.
Bigger, not better. Tripled PR sizes still have to be reviewed. The staff engineer quoted above says human reviewers catch changes “that are not needed at all, or those that are 10 times bigger than they should be,” and sees “accelerated slop produced all around me.”
The verification tax. DORA models a J-curve: a productivity dip while teams learn, review AI code and adapt testing and change approval to higher volumes. Its sample calculation books a $344,000 downtime cost because the change failure rate rises from 5% to 6%.
Security debt. Veracode’s 2026 GenAI Code Security Report finds that models write syntactically correct code nearly every time, but the average security pass rate is 56%, barely up from 55% in its first report. Cross-site scripting tasks passed just 15% of the time. Veracode sells security testing, so treat this as a vendor benchmark. The direction still fits: writing code got cheap, checking it did not.
Attackers are speeding up too. On October 22, The Pragmatic Engineer’s The Pulse opened with a series of high-profile incidents and breaches in a single day, “many fueled by AI” (paywalled; I read only the summary). I covered that week in 72 hours of security headlines.
How to measure it
Infographic 03 · How to measure it
Measure outcomes, not consumption.
Token spend is an input. What matters is what ships, stays shipped and stays secure.
Don’t steer by
activity
- Token consumption and leaderboards (“tokenmaxxing”)
- Raw PR counts
- Lines of code
- Self-reported speed
Track instead
outcomes
- Cost per accepted change
- Rework and change failure rate
- Stories closed vs. stories coming in
- Security findings per change
- Lead time and incident resolution
Budget for
hidden costs
- The J-curve: a dip while teams learn and adapt
- The verification tax: reviewing AI code
- Instability: DORA’s example books $344K when change failures rise from 5% to 6%
- Security: 44% of AI coding tasks produced a known vulnerability in Veracode’s test
DORA’s worked example
Illustrative, not measured
- 500-person engineering org, $176,000 fully loaded cost per head
- Invest about $8.4M, return about $11.6M
- ROI = (value − investment) ÷ investment ≈ 39% in year one, payback ≈ 8 months
Guardrails that work
Examples DORA cites
- Shopify: usage dashboard instead of a leaderboard, circuit breakers for runaway agents and spend spikes
- Starburst: ignores token maximization, measures standard DORA metrics, velocity and incident resolution
“Treat these calculations as a high-uncertainty estimate meant to spark a conversation.”DORA, ROI of AI-assisted Software Development (2026.01), as quoted by InfoQ
Sources: DORA ROI of AI-assisted Software Development (2026.01) via InfoQ (11 May 2026); DORA on tokenmaxxing (2 Jun 2026); Veracode 2026 GenAI Code Security Report (28 Jul 2026). Scorecard selection: Thorsten Meyer.
Powered by Thorsten Meyer AI
DORA’s formula is plain: value minus investment, divided by investment. Its worked example, a 500-person engineering organization at $176,000 fully loaded cost per head, invests $8.4 million, returns about $11.6 million and reaches a 39% first-year ROI. DORA calls it “a high-uncertainty estimate meant to spark a conversation.” Use the structure, not the number.
What I would track:
- Cost per accepted change: AI spend divided by changes that pass review and stay in production.
- Rework and change failure rate, to catch the instability tax early.
- Stories closed versus stories coming in, the closest everyday proxy for business value.
- Security findings per change in review and scanning.
- Spend with guardrails. DORA points to Shopify, which replaced its AI leaderboard with a usage dashboard and circuit breakers for runaway agents and spend spikes.
What this means for companies
This section and the next are my assessment.
- Budget a range, not a line. Plan a usage-based scenario for 2027: subsidies are ending, and Copilot already bills by the token.
- Fund verification as seriously as generation. Review, testing and change approval are where the gains disappear.
- Expect the J-curve, and measure it.
- Don’t reward consumption. Microsoft found that first use spread mainly through social networks. Visible peer use beats leaderboards.
- Put security in the AI budget. Scanning and review are part of what AI coding costs. Formal methods may change that, as Claude showed by turning fuzz tests into proofs, but not yet for most stacks.
What this means for users
The job is shifting from typing code to judging it.
- Review is the job. In the Stack Overflow Developer Survey released October 6, 48% of respondents trust AI output when they can easily verify it; only 6.6% trust it for important work decisions.
- Protect your learning. In an Anthropic randomized experiment with 52 developers learning a new Python library, the AI group scored 17% lower on a comprehension quiz, without a significant speed gain on average. Those who stayed engaged kept learning; those who fully delegated did not.
- Your spend is becoming visible. Caps, a Slack channel that names top spenders, and an AI skill that audits your usage before any budget increase already exist.
- Leverage favors small, high-agency teams. Freelancers and sole proprietors rose to 11% of Stack Overflow respondents, from 4% in 2025, and the clearest winners in Orosz’s data are small teams that know what to build.
Where the evidence stops
- The Pragmatic Engineer data are interviews with engineers who chose to answer, mostly anonymized and partly paywalled: case studies, not a sample.
- CodeRabbit, Gitar and Veracode are vendors. Pull requests are a proxy, as Microsoft says explicitly.
- METR’s speedup estimates are not statistically significant, and METR treats them as a lower bound.
- DORA’s ROI figures are an illustrative model, not measured outcomes.
- The skill study used GPT-4o in a 35-minute task; today’s agents work differently.
Verdict
AI coding pays where work is verifiable, repetitive or exploratory, and where a team can absorb more change. It burns money where tokens become a target and review becomes the bottleneck. The companies that can answer the ROI question measured what they shipped before AI and kept measuring afterwards. If you can’t say that today, start this quarter, before the 2027 budget is locked and the subsidies run out.
Sources
- The Pragmatic Engineer (Gergely Orosz): What’s the ROI of software engineers’ spending on AI? (October 20, 2026; freely readable part only, includes CodeRabbit and Gitar data)
- The Pragmatic Engineer: The Pulse: Time to rethink software security because of AI? (October 22, 2026; summary only, paywalled)
- Murphy-Hill, Butler, Savelieva: Adoption and Impact of Command-Line AI Coding Agents: A Study of Microsoft’s Early 2026 Rollout of Claude Code and GitHub Copilot CLI (arXiv, July 1, 2026)
- METR: We are Changing our Developer Productivity Experiment Design (February 24, 2026) and Measuring the Self-Reported Impact of Early-2026 AI on Technical Worker Productivity (May 11, 2026)
- DORA (Google Cloud): ROI of AI-assisted Software Development report (2026.01, page updated April 22, 2026), figures as summarized by InfoQ (May 11, 2026); Finding balance in the era of tokenmaxxing (June 2, 2026)
- GitHub: GitHub Copilot is moving to usage-based billing (April 27, 2026)
- Veracode: 2026 GenAI Code Security Report (July 28, 2026)
- Stack Overflow: 2026 Developer Survey press release (October 6, 2026) and trust data
- Shen and Tamkin (Anthropic): How AI Impacts Skill Formation (arXiv, January 28, 2026)
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
