AI Coding Tools and Your Dev Quote in 2026: What the Data Actually Shows

A vendor tells you their team uses AI coding tools, so your project will cost less and ship faster than it would otherwise. That pitch shows up in most quotes you'll see in 2026, and it rarely comes with anything beyond the vendor's own say-so. The actual 2025-2026 research on AI coding tools and software development cost tells a messier story. In a controlled study, AI assistance slowed experienced developers down. Pull requests with AI involved sat longer before anyone reviewed them. Reviewers flagged more issues in AI-coauthored code than in human-only code. And the token bill for running these tools is turning into a cost line serious enough that Gartner is now forecasting it against developer salaries. None of this means AI tools are useless on a real project. It means "we use AI" is not, on its own, a reason to accept a lower price or a shorter timeline, and it's no substitute for understanding the hidden costs buried in a typical development quote.

What the 2025-2026 research actually found

Four separate data sets came out in the back half of 2025 and the first half of 2026, from four different organizations with four different incentives. None of them agree with the "AI makes it cheaper and faster, full stop" pitch.

Developers got slower, not faster

METR's July 2025 study put the productivity claim to a real test. Sixteen experienced open-source developers, each averaging around five years on their own projects, worked through 246 real GitHub issues. Each task was randomly assigned to allow or disallow early-2025 AI tools, mostly Cursor Pro paired with Claude 3.5 or 3.7 Sonnet.

The developers finished 19% slower when they used AI tools than when they didn't. Before the study started, they predicted AI would make them 24% faster. After finishing, most still believed they'd been roughly 20% faster, even though the clock said otherwise. That gap between felt speed and measured speed is the whole problem with taking a vendor's word for it: the people doing the work can be confident and wrong at the same time.

METR itself frames this as a snapshot of one set of tools at one point in early 2025, not a permanent verdict on AI coding assistance. Models and workflows have moved since. But "it'll obviously be faster" was never a safe assumption, and this is the closest thing to a controlled test of that assumption that exists right now.

Pull requests sit longer before anyone picks them up

LinearB's analysis of more than 8.1 million pull requests across 4,800-plus organizations, covering all of 2025, found that pull requests touched by agentic AI tools took 5.3 times longer to get picked up for review than pull requests without AI involvement. They also merged less reliably: 32.7% of AI-assisted pull requests closed within 30 days, versus 84.4% of human-only ones.

A PR that takes longer to land isn't automatically a bad PR. But "pickup time 5.3x longer" and "merge rate 13 points lower" are not the hallmarks of a faster pipeline, and a vendor who tells you AI speeds up delivery is making a claim this data doesn't back up, at least not once you look past the moment the code gets written and into the review queue where it actually has to get approved.

Reviewers flag more issues in AI-coauthored code

CodeRabbit reviewed 470 open-source pull requests, 320 of them AI-coauthored and 150 human-only, and counted what its reviewers flagged in each. AI-coauthored pull requests averaged 10.83 code-review issues per PR, against 6.45 for human-only ones, roughly 1.7 times as many. Security issues showed up 1.5 to 2 times more often in the AI-coauthored set, readability issues more than three times as often, and the highest-severity issues around 1.4 to 1.7 times more often depending on category.

Worth saying plainly: CodeRabbit sells AI-focused code review tooling, so a finding that AI code needs more reviewing is good for its business. The measure here is code-review issues per PR, not an independently audited defect rate, and it isn't proof that AI-written code ships more bugs to production. It is evidence that whoever reviews AI-coauthored pull requests at your vendor has more to catch, which has its own cost even if every issue gets caught before release.

The token bill is becoming its own cost line

Gartner's June 2026 forecast predicts that AI coding tool costs will surpass the average developer's salary by 2028, driven by rising token consumption rather than by seat licenses. The same survey found that roughly a quarter of tech leaders already spend $200 to $500 per developer per month on AI coding tokens, and 6% spend more than $2,000 per developer per month.

That's not a cost that shrinks a quote. It's a cost that has to live somewhere, either absorbed by the vendor, passed to you as a line item, or quietly built into the hourly rate you were quoted in the first place.

Study	Finding	What it measured	_url
METR (July 2025)	AI tools made developers 19% slower, despite predicting 24% faster	Randomized controlled trial, 16 developers, 246 real GitHub issues	METR's randomized trial
LinearB (2026 report, FY2025 data)	Agentic-AI pull requests took 5.3x longer to get picked up; merged within 30 days at 32.7% vs. 84.4%	8.1M+ pull requests, 4,800+ organizations	LinearB's 2026 benchmarks report
CodeRabbit (December 2025)	AI-coauthored PRs averaged 10.83 review issues vs. 6.45 for human-only, about 1.7x	470 open-source pull requests (320 AI-coauthored, 150 human-only)	CodeRabbit's code-generation whitepaper
Gartner (June 2026)	AI coding costs predicted to surpass average developer salary by 2028	Survey + forecast; roughly a quarter of tech leaders spend $200-500/developer/month on tokens, 6% spend $2,000+	Gartner's token-cost prediction

The "46% of code is AI-written" line is three years out of date

You'll still hear a vendor cite a number like "46% of all code in active repositories is AI-written" as proof that AI has already taken over development. That framing misstates what was actually measured. The original claim traces to GitHub's own 2022-2023 Copilot-for-Business figures, publicly cited by CEO Thomas Dohmke: 46% of code written in Copilot-enabled files was accepted from Copilot's own suggestions, a suggestion-acceptance rate among GitHub Copilot users specifically, measured three to four years ago.

Recent "AI statistics" roundups dropped both qualifiers, the "Copilot-enabled files" part and the 2022-2023 date, and now present the number as a current, universal measure of how much code everywhere is AI-written. It isn't that. It's a three-year-old acceptance rate from one tool's own users, repackaged to sound like a 2026 fact about the entire industry. If a vendor leads with this statistic unprompted, ask them where they got it. There's a decent chance they can't tell you, because it's being recirculated secondhand at this point, stripped of the context that made it meaningful in the first place.

Questions to ask before you sign a quote that leans on "we use AI"

A "we use AI" line in a proposal isn't a red flag by itself. Plenty of legitimate teams use these tools well. But the data above gives you concrete questions to ask before the discount or the shortened timeline gets written into a contract.

Which parts of the work are the AI tools actually speeding up? Code generation, test writing, and documentation are different from code review, debugging, and the architecture decisions that determine whether the thing scales. A vendor who can name the specific tasks is telling you something real. One who answers in generalities is reciting a pitch.

Who reviews the AI-generated code, and how long does that take? If review is where AI-assisted work tends to slow down, as LinearB's pickup-time data and CodeRabbit's issue counts both suggest, a vendor's timeline needs to account for that stage, not just the stage where code gets typed.

Is the lower price based on their own measured output, or on industry marketing? Ask for their actual numbers: hours per feature before and after they adopted these tools, on comparable past work. A team that's tracked this will have an answer. A team repeating an unsourced "AI makes us faster" line is quoting someone else's press release, not their own data.

What happens to cost if token usage runs higher than expected? Given where Gartner says token spend is headed, a fixed quote that doesn't address this is either absorbing the risk (ask how) or passing it to you later as a change order.

These questions sit inside the broader list of red flags in a vendor pitch worth checking before you sign anything, and they matter more now that "AI" has become a line item vendors expect you not to question.

How Appsoluty prices instead

We don't quote a lower rate because of which tools our developers use, and we don't build "AI acceleration" into a timeline we can't back up. Our rate is $70 an hour. Before work starts, we estimate the hours a task will take and show you that projected cost upfront, so you see the number before you commit to it, not after. Depending on the project, you engage us on a milestone basis or an hourly basis, billed weekly for hours actually worked, and work begins once the contract is signed.

That's not a discount pitch. It's the opposite of one: a rate you can check against what a quoted hourly rate actually buys elsewhere in the market, and hours you can verify against the work delivered, rather than a number justified by a tool claim you can't audit. If you want to see how that compares once you factor in everything a quote doesn't show on its face, the real total-cost-of-ownership math is the place to run it.