2026 年 AI 編碼生產力分析:開發成本下降,決策成本未變
AI Coding Productivity in 2026: Building Got Cheap, Deciding Didn't
AI 編碼工具令軟件開發更普及、部分任務更快,但不同研究的提速結果不一,未顯示成熟團隊能大幅加快交付合適產品。2023 年一項試驗中,使用 Copilot 的開發者完成新 HTTP 伺服器快 55.8%;METR 2025 年試驗則發現,資深開發者在大型成熟程式碼庫中使用 AI 慢 19%。
AI tools made building software cheaper and more common. The evidence on speed is mixed, and the cost of building the wrong thing has not moved. Here is what the 2025-2026 data actually says.
AI coding productivity rose in reach more than in proven speed. In 2025, 84% of developers used or planned to use AI tools, yet METR's controlled trials found experienced developers 19% slower in early 2025 and only weak evidence of speed-up by late 2025. Meanwhile 43% of failed startups still cite poor product-market fit. Building got cheaper; knowing what to build did not.
Key takeaways
- AI coding is now the default. 84% of developers use or plan to use AI tools, and 50.6% of professional developers use them daily (Stack Overflow, 2025). Nearly 80% of new GitHub developers use Copilot in their first week (GitHub Octoverse, 2025).
- Speed gains depend on the task. Developers finished a greenfield HTTP-server task 55.8% faster with Copilot in a 2023 trial, but experienced developers on large, mature codebases were 19% slower with AI in METR's early-2025 trial.
- People overestimate the gain. METR's developers expected a 24% speed-up and, after being slowed down, still believed AI had made them 20% faster.
- The expensive half hasn't moved. In CB Insights' 2026 analysis of 431 VC-backed shutdowns, 43% cited poor product-market fit (CB Insights, 2026).
- Most features still go unused. Pendo found 80% of features in the average product are rarely or never used (Pendo, 2019).
How many developers use AI coding tools in 2026?
Almost all of them, in some form. The most recent large surveys agree on reach even where they disagree on impact.
What the Stack Overflow Developer Survey 2025 found
The 2025 survey, published in July 2025, reports that 84% of respondents use or plan to use AI tools in development, from 33,662 responses. Among professional developers, 50.6% use them daily.
Agents are earlier on the curve: 14.1% of respondents use AI agents at work daily, 9% weekly, and 37.9% do not plan to use them at all (Stack Overflow, 2025).
What GitHub Octoverse 2025 found
GitHub's Octoverse report, published on 28 October 2025, counts 180 million-plus developers on the platform, with more than 36 million joining in 2025. Activity grew with them:
- Nearly 1 billion commits pushed in 2025, up 25% year on year.
- 518.7 million pull requests merged, up 29%.
- More than 1.1 million public repositories import an LLM SDK, up 178%.
More code is being shipped by more people. That is what "building got cheap" means in practice (GitHub, 2025). It also means more products competing for the same buyers, increasingly inside AI answers; see product explainability.
Does AI actually make developers faster?
Sometimes, by a lot; sometimes not at all. The answer depends on the task, the codebase and how you measure.
| Study | Who and what | Result |
|---|---|---|
| Peng et al., 2023 (controlled trial) | Recruited developers writing an HTTP server in JavaScript | Copilot group finished 55.8% faster |
| METR, early 2025 (randomised trial) | 16 experienced open-source developers, 246 real tasks in repositories averaging 1M+ lines of code | 19% slower with AI allowed (95% CI +2% to +39%) |
| METR, late 2025 update (published Feb 2026) | 57 developers, 800+ tasks; 10 returning, 47 new | Returning developers −18% time (CI −38% to +9%); new developers −4% (CI −15% to +9%) |
Sources: Peng et al., 2023; METR, 2025; METR, 2026.
Why do the studies disagree?
- Greenfield vs mature code. Writing a small new server from scratch is where AI shines. Changing a million-line codebase you know well leaves less room for help and more room for review overhead.
- Perception vs measurement. METR's early-2025 developers forecast a 24% speed-up and afterwards believed they had been 20% faster, while the clock said 19% slower. Self-reported productivity is not evidence.
- Selection effects. In the late-2025 update, METR reports that 30% to 50% of developers chose not to submit some tasks, and that developers unwilling to work without AI declined to take part. METR says the true speed-up could be much higher for the tasks selected out, and calls its own data only very weak evidence of an increase since early 2025.
What about trust and quality?
In the Stack Overflow survey, 46% of developers distrust the accuracy of AI output and only 33% trust it. The top frustration, cited by 66%, is answers that are "almost right, but not quite". Faster typing does not help if review and rework eat the gain.
The honest summary on AI coding productivity: AI makes producing code cheaper and more accessible, and often faster for new or well-bounded work. It has not been shown to make mature teams dramatically faster at shipping the right thing.
If building is cheaper, why do products still fail?
Because the bottleneck was never only typing speed.
Startups fail on demand, not code
CB Insights' March 2026 analysis looked at 431 VC-backed companies that shut down since 2023 and identified reasons for 385 of them:
- 70% ran out of capital, which CB Insights calls almost always the final cause rather than the root problem.
- 43% had poor product-market fit.
- 29% cited bad timing or macro conditions.
- 19% had unsustainable unit economics.
Companies cite several reasons, so these add up to more than 100% (CB Insights, 2026). Cheaper code makes the runway last longer per feature. It does nothing for a product nobody needs.
Most shipped features are barely used
Pendo's 2019 Feature Adoption Report analysed 615 Pendo subscriptions over three months. It found that 80% of features in the average product are rarely or never used, and that 12% of features generate 80% of daily usage volume. Pendo extrapolated that publicly traded cloud software companies invested up to $29.5 billion building those features (Pendo, 2019). We trace that statistic in detail in why 80% of features are rarely or never used.
Product teams feel the risk
Atlassian's State of Product 2026 surveyed more than 1,000 product professionals in the US and Europe and found that 84% of PMs fear their products will fail (Atlassian, 2026). Mind the Product's write-up of the same report notes that 80% of teams still don't involve engineers during ideation, problem definition or roadmap creation (Mind the Product, 2026).
AI coding productivity and product failure: the numbers in one place
| What | Number | Source and year |
|---|---|---|
| Developers using or planning to use AI tools | 84% | Stack Overflow, 2025 |
| Professional developers using AI daily | 50.6% | Stack Overflow, 2025 |
| Developers who distrust AI output accuracy | 46% | Stack Overflow, 2025 |
| New GitHub developers using Copilot in week one | Nearly 80% | GitHub Octoverse, 2025 |
| Pull requests merged on GitHub in a year | 518.7 million (+29%) | GitHub Octoverse, 2025 |
| Speed-up on a greenfield coding task | 55.8% faster | Peng et al., 2023 |
| Effect on experienced developers, mature repos | 19% slower | METR, 2025 |
| Failed VC-backed startups citing poor PMF | 43% | CB Insights, 2026 |
| Features rarely or never used | 80% | Pendo, 2019 |
| PMs who fear their product will fail | 84% | Atlassian, 2026 |
What should teams build when code is cheap? A 6-step method
When a feature costs days instead of weeks, the risk shifts from "can we build it?" to "should we?". Six steps keep cheap building from becoming cheap waste:
- Write the problem before the solution. One paragraph: who has the problem, how often, and what it costs them. Use our product problem statement template if you need a format.
- Count the evidence. List the calls, tickets and usage data behind the problem. Fewer than 3 independent sources means you are guessing.
- Run the cheapest test first. A fake door, a concierge test or five interviews often settles the question faster than a build.
- Prototype as a test, not a commitment. AI lets you build a working prototype in hours; treat it as a way to watch real users, then decide. Our 7-day plan for validating an AI prototype covers this step.
- Set a kill criterion before launch. For example: "If fewer than 20% of active accounts use it within 30 days, we remove it."
- Check adoption after launch. Most teams never look again, which is how the 80% accumulates.
The argument that "we can build it in an hour, so just build it" ignores maintenance, support and product complexity; we take it apart in how to say no to "we can build this in 1 hour with AI".
Worked example: what does a wrong bet cost now?
Illustrative scenario: a 6-person startup ships 10 features a quarter. AI tools cut build time per feature from 2 engineer-weeks to 4 engineer-days, so the quarter's build effort falls from 20 engineer-weeks to 8.
If the Pendo pattern holds and 8 of those 10 features are rarely used, the team still spends 6.4 engineer-weeks building features nobody needs. It also keeps paying to maintain, document and support all 8. Cheaper building lowered the price of each wrong bet; it did not lower the number of wrong bets.
Now suppose the team spends 1 extra day per feature on evidence (steps 1–3 above) and cuts 4 of the 8 weak ideas before building. That is 10 days of discovery to save 16 engineer-days of build, plus the ongoing upkeep of 4 features. The return comes from better choices, not faster code.
What doesn't this data show?
- Pendo's figure is from 2019 and from Pendo customers. It is the best-known measurement, not a law of nature.
- CB Insights covers VC-backed shutdowns only. Bootstrapped and still-alive companies are not in the sample.
- METR's trials are small (16 and 57 developers) and focus on open-source work. METR itself warns against reading its central estimate as the general effect of AI.
- Survey samples self-select. Stack Overflow respondents and GitHub users are not every developer.
- None of these studies links AI coding to feature adoption. The claim that cheaper building increases waste is an inference from the numbers, not a measured result.
Where Specky fits
Specky works on the expensive half. It reads customer calls, feedback, Slack, Jira and analytics into a Product Graph, drafts specs with the customer evidence cited inline, and checks after launch whether the bet worked. If you are choosing tools for that job, our comparison of AI product management tools covers Specky and seven alternatives, including where others fit better. As AI makes output cheap, pricing is shifting too; see per-seat pricing is dying.
Frequently asked questions
Does AI make developers more productive?
It depends on the task. A 2023 controlled trial found developers using Copilot finished a new HTTP server 55.8% faster. METR's 2025 trial found experienced developers working in their own large codebases were 19% slower with AI. METR's late-2025 follow-up pointed towards speed-up, but its confidence intervals included zero, so the evidence for a large general gain is still weak.
What percentage of developers use AI coding tools?
In the Stack Overflow Developer Survey 2025, 84% of respondents used or planned to use AI tools in their development process, and 50.6% of professional developers used them daily. GitHub's Octoverse 2025 reports that nearly 80% of new developers on GitHub use Copilot within their first week.
What did the METR study on AI and developer productivity find?
METR's early-2025 randomised trial gave 16 experienced open-source developers 246 real tasks. With AI allowed, tasks took 19% longer, even though developers expected a 24% speed-up and believed afterwards they had been 20% faster. A February 2026 update with 57 developers found point estimates of 4% to 18% speed-up, with wide confidence intervals.
Why do startups fail in 2026?
CB Insights' March 2026 analysis of 431 VC-backed shutdowns found that 70% ran out of capital, usually the final cause rather than the root one. Among underlying reasons, 43% cited poor product-market fit, 29% bad timing or macro conditions, and 19% unsustainable unit economics. Many companies cited more than one reason.
Is it true that 80% of software features are rarely used?
That figure comes from Pendo's 2019 Feature Adoption Report, which analysed 615 Pendo subscriptions over three months and found 80% of features in the average product are rarely or never used. It is a real measurement but dated and limited to Pendo customers, so treat it as a strong signal rather than a universal constant.
Sources
- Stack Overflow — 2025 Developer Survey: AI
- GitHub — Octoverse: A new developer joins GitHub every second as AI leads TypeScript to #1 (2025)
- arXiv — Peng, Kalliamvakou, Cihon and Demirer, The Impact of AI on Developer Productivity: Evidence from GitHub Copilot (2023)
- METR — Early-2025 AI and experienced open-source developer productivity study (2025)
- METR — February 2026 update on developer productivity with late-2025 AI tools
- CB Insights — Top reasons startups fail (2026)
- Pendo — The 2019 Feature Adoption Report
- Atlassian — State of Product 2026
- Mind the Product — Four things we learned from Atlassian's State of Product 2026
This story was originally published on 17 July 2026 at https://www.specky.space/blog/building-got-cheap-what-to-build-2026-data.
來源:DEV Community · dev.to