In July 2025, a research group called METR paid 16 experienced open-source developers to work through 246 real tasks in codebases they already knew well. Half the tasks allowed AI tools. Half didn’t. Before starting, the developers predicted AI would make them 24% faster. It made them 19% slower. Afterwards, having just done the work, they still believed it had sped them up by 20%.
Hold the slowdown lightly. METR has since run the experiment again with late-2025 agentic tools and got an 18% speed-up for the same developers, then said in February 2026 that the result is “only very weak evidence” because so many developers now refuse to work without AI that the sample is bent. Hold the perception gap tightly, though. That part has never wobbled. People are bad at estimating how much faster AI made them, and your development quotes are being priced by people making exactly that estimate.
The cost of producing code is falling. The cost of everything wrapped around code isn’t falling, and some of it is rising because there’s now more code to wrap. Your build budget will come down. It will come down by less than the demo suggested, and from a narrower set of line items than anyone selling you the saving has mentioned.
The short version
- 81% more duplicated code, and 45% of AI coding tasks fail security. You inherit both.
- Only 3 lines on a typical quote should fall. Most of the rest should rise.
- A fuzzy brief now costs around $10,000 in wasted weeks, not a clarifying question.
- 3 questions that tell you if a quote’s saving is real or cosmetic.
The output went up. The clean-up went with it.
Google’s 2025 DORA report surveyed nearly 5,000 technology professionals. 90% use AI at work and more than 80% say it’s made them more productive. In the same survey, 30% report little or no trust in the code the AI writes, and DORA found AI adoption still carries a negative relationship with software delivery stability. More gets shipped. More breaks.
Stack Overflow’s 2025 developer survey puts a shape on why. 46% of respondents actively distrust the accuracy of AI output and 3% highly trust it. The single biggest frustration, cited by 66%, is output that’s almost right but not quite. The second, at 45%, is that debugging AI-written code takes longer.
GitClear tracked 623 million code changes between 2023 and 2026. Duplicated code blocks are up 81% and at the highest level on record. Refactoring, the work of tidying up rather than adding, fell from 21% of changed lines in 2022 to 3.8% in 2026. Copy-paste inside a single commit rose from 9.4% to 15.7%. Duplication is cheap to create and expensive to own, which is a decent one-line summary of the whole problem.
The bottleneck moved to review. LinearB analysed 8.1 million pull requests from 4,800 organisations across 42 countries for its 2026 benchmarks and found AI-generated changes were accepted at 32.7%, against 84.4% for hand-written ones. AI-authored work also waits 4.6 times longer before anyone picks it up to review.
Then security. Veracode has tested more than 150 models across 80 coding tasks in 4 languages. As of its Spring 2026 update, 55% of generation tasks produce secure code, meaning 45% ship a known weakness. Cross-site scripting passes 15% of the time. Log injection, 13%. Barely moved in 2 years.
None of this says don’t use the tools. I use them daily, on code I’ll be the one maintaining. It says the saving lands in one place and the bill in another.
What actually falls, and what doesn’t
| Line item | Direction | Why |
|---|---|---|
| Writing the first version | Down hard | This is the part the tools are genuinely good at |
| Boilerplate, scaffolding, admin screens | Down hard | Well-solved, well-represented in training data |
| Prototypes and throwaway builds | Down hard | Nobody has to maintain them |
| Scoping and specification | Flat or up | Vague briefs now produce 3,000 lines of confidently wrong code instead of a question |
| Code review | Up | More volume, lower trust, longer queues |
| Integration and testing | Up | Same reason |
| Security review | Up | 45% failure rate on a much larger denominator |
| Maintenance in year 2 | Up | Duplication is up 81% and refactoring is down to 3.8% |
Worldwide IT spending is forecast to grow 13.5% in 2026 to US$6.31 trillion, according to Gartner’s May 2026 revision. That isn’t the shape of an industry getting cheaper. It’s the shape of one where the unit cost dropped and the unit count exploded.
What this means if you’re buying, not building
If you run a business with real revenue and no developers of your own, 3 things follow.
Stop paying full rate for the first version. Prototypes, internal tools, admin panels, one-off scripts, the fourth landing page variant. This work has genuinely collapsed in price. If a quote for it looks like a 2023 quote, it is one. Ask what changed.
Start paying properly for the specification. The most expensive hour in a build has always been the one where nobody wrote down what the thing was meant to do. That hour is now worse, because a vague brief no longer stalls. It produces something plausible that fails in month 4. In Australia, SEEK’s job-ad data puts the average advertised salary for software engineer roles in ICT at about $128,000. Four wasted weeks of that is roughly $10,000, and it will be spent by someone whose job is to build what you asked for rather than what you meant.
Ask who reviews it and how long that takes. If the answer is thin, you’re not buying a cheaper build. You’re buying the same build with the quality-control step removed, and taking delivery of the difference as maintenance. Once the work is under way, 6 questions sent by email will tell you whether the invoices are fair.
The 3 questions I’d put on any development quote from here. What portion of this is generation and what portion is review? Who signs off on security, and when? What does year 2 cost? If the person quoting can’t separate those, they haven’t repriced their work. They’ve just repriced their typing.
Code got cheap. Judgement didn’t, and there’s now a great deal more code that needs some.