AI Coding Tools Add 180 Percent More Code, Not Users

AI coding tools raised commit activity by 180% for more than 100,000 GitHub developers, a May 2026 NBER paper finds. Software releases rose only 30%, and usage on major app stores did not rise with the extra supply.

That split is an old one. Whenever making software got cheaper, catalogs filled faster than anyone opened them. Blogs did it. Steam did it. The phone stores did it after 2008. The new paper shows the same curve, now drawn by autocomplete, sync agents, and async agents.

What the 180 Percent Coding Gain Ships

Mert Demirer of MIT Sloan, Leon Musolff of Wharton, and Liyuan Yang of Boston University tracked developers from 2022 through 2026 in a matched event study of 100,000 GitHub developers. They folded in Microsoft telemetry on GitHub Copilot from April to December 2025, plus public traces of Claude Code and Codex, and compared each adopter with a similar developer from the same calendar week a year earlier.

Autocomplete, the first Copilot-style line-suggester, raised commits about 40%. Adding sync agents that edit beside the developer, including Claude Code run locally, took the cumulative gain to 140%. Async agents that take a prompt and come back with a pull request, including Copilot’s GitHub agent and Codex, pushed it to 180%.

Those commits do not survive the rest of the production chain. The 180% coding gain shrinks to 50% more projects and 30% more releases. Autocomplete alone raised lines of code 228% and releases 10%. Sync agents still only lifted releases 20%.

HOW THE GAINS FADE DOWN THE CHAIN

The authors map that fade to a weak-link model. They estimate an elasticity of substitution of 0.25 between AI effort and human effort, which means the two do not swap cleanly. Extra code still needs someone to review it, merge it, and ship it. GitHub itself is huge in the background, with 180 million developers and 395 million public repositories in the 2025 Octoverse figures the paper cites, so the sample is not a toy corner of the industry.

Demirer has been hearing the task-level story in shops already. “We’re hearing they can do in a matter of minutes what used to take an entire day,” he said. The NBER results say that minute-scale coding is real, and that it is the wrong layer to watch if the question is finished software.

The Same Glut Already Hit Games and App Stores

Steam ran a quieter version of this experiment for years. The Steam Revenue Report compiled by Totally Human Media and VaporLens, using review counts to estimate sales, put median lifetime earnings of $3,499 across 100,720 paid games with at least one review. The top 1% of those games take 84.5% of estimated revenue. Another 8,290 paid games have zero reviews, sitting below the floor the method can even price. Cheap tools to ship a game did not cheapen attention.

Phone software is now on the same path. Similarweb says monthly app releases have jumped 50% year over year since early 2025, to around 120,000 a month across the big stores. On Apple’s App Store, the NBER authors’ marketplace panel shows monthly new releases rising from around 30,000 before coding agents arrived in early 2025 to roughly 100,000 by April 2026. That is more than three times the earlier flow. Total usage, measured in the first three months after launch, did not rise on Apple, Google Play, the Chrome Web Store, or SourceForge.

THREE MARKETS WHERE SUPPLY OUTRAN USE

• Steam’s paid catalog: Half of 100,720 measured games earned under $3,499 over their whole life, and the top 1% took 84.5% of estimated revenue.
• New Android apps: Similarweb finds that 75% of new Android titles never reach 1,000 downloads, even as monthly storewide releases sit near 120,000.
• Apple’s App Store: New monthly releases moved from around 30,000 to roughly 100,000 by April 2026, while early usage across four stores stayed flat or fell.

Investor Andrew Chen wrote that split down in March 2025, before this paper existed. If vibe coding makes software trivial to build, he said, the bottlenecks shift to distribution and network effects, and the first prompt-built product is not the one that wins. The NBER usage charts are that bet showing up in store data.

Overnight shipping is still the flex on developer timelines. People post a gallery app built in one night, or an agent that wires a pipeline and patches it before anything is saved. That culture keeps feeding stores. It does not create an audience. A lot of the new titles were never aimed at 50,000 users; they are scratch tools for a few dozen people. Public stores still have to index them, rank them, and bury them.

Review Queues Catch What Agents Cannot Merge

Inside a repo, the choke is not typing. Modern work still runs through a six-step chain the MIT Sloan write-up lays out: lines of code, files, commits, pull requests, merged projects, and releases. Agents now sit on the first four. People still own the last two.

WHERE THE HUMAN HAND STILL SITS

• Review: Someone has to read the diff, including the agent-authored commits the 180% figure counts.
• Integration: Changes still have to fit tests, style, and the rest of the tree before a merge.
• Release: Async agents, by design, cannot ship the binary; the paper does not even score them on that last step.

Musolff is blunt about the next product idea, which is to let the same model grade its own work.

If the same AI that wrote the code also reviews it, that doesn’t really solve the problem. The review just isn’t of the same quality.

Leon Musolff, Wharton economist

Some shops are already building reviewer bots anyway. The 0.25 elasticity is the authors’ way of saying that extra machine output still wants extra human time downstream. Make more pull requests, and the queue gets longer unless review is automated with real judgment, which Musolff does not think the writer model can supply.

That is also why a randomized trial from METR, whose authors the NBER paper thanks, can sit beside a 180% commit boom without fully contradicting it. In early 2025, 16 experienced open-source developers took 19 percent longer on real issues in their own large repos when they were allowed to use Cursor Pro and Claude 3.5/3.7 Sonnet. They had forecast a 24% speedup, and after the work they still believed they had been 20% faster. Commits can rise while hard, familiar tasks get slower, if people accept more machine drafts and then spend the afternoon cleaning them.

Why New Apps Still Find Almost No Users

On the four stores in the paper, new listings rose from mid-2025, hardest on the Apple App Store and the Chrome Web Store, milder on Google Play, and barely at all on SourceForge. Usage in the first three months after launch did not rise on any of them. The share of new titles that never reach even a modest audience went up, which is the glut showing up as dead listings rather than as more time in the apps people already had.

Similarweb’s public snapshot of the same wave is harsher on the tail. Since early 2025, monthly releases are up 50%, and 75 percent never reach 1,000 downloads on Android. Discovery still leans toward names users already know. Building dropped in cost. Being found did not.

WHAT WE KNOW

• Supply: New applications rose across the four stores after mid-2025, with Apple moving from around 30,000 monthly releases to roughly 100,000 by April 2026.
• Early usage: Total engagement in the first three months after launch did not increase on any of the four marketplaces.
• The tail: The share of new apps that fail to reach a modest audience has risen, and Similarweb’s Android cut puts 75% under 1,000 downloads.

WHAT IS UNCONFIRMED

• Quality versus discovery: The authors cannot yet tell whether the extra apps are worse, or whether people simply cannot find the good ones in the flood.
• Enterprise software: Internal tools never hit these stores, so a company could still be shipping more useful software that this panel never sees.

Musolff offers both readings. It may be harder to discover anything in a flood. Or shipping may still leave a second skill on the table, iterating with users after the binary is live. Either way, the consumer side is now a second weak link, sitting after the review queue.

Humans Still Merge What Agents Open

The tools themselves have been walking down that chain, which is why the authors do not treat the 30% as a ceiling. Autocomplete only helped at the cursor. Sync agents write and edit in the working tree. Async agents open the pull request, run tests on a cloud machine, and wait for a human merge. Each generation takes one more layer that used to be a person.

HOW THE TOOLS MOVED DOWN THE CHAIN

1. 2022: GitHub Copilot autocomplete becomes the first widely measured AI coding layer; the paper’s adopters show a 40% commit gain and a 10% release gain.
2. March 2023: Cursor joins a field that already includes Copilot chat in the editor; the NBER study still does not treat chatbots as their own generation.
3. Early 2025: Agentic tools become easy to assign a whole task; new app listings inflect upward on the stores.
4. April 2026: Apple’s monthly new releases sit near 100,000, while first-three-month usage on the four stores has still not risen.
5. May 2026: NBER Working Paper 35275 reports the 180/50/30 fade and the 0.25 elasticity.

An earlier Copilot experiment, also involving Demirer and Musolff, had already shown a smaller, cleaner gain at the pull-request layer. At Microsoft, 1,663 developers completed 12.92% to 21.83% more pull requests a week with Copilot. At Accenture, 311 developers gained 7.51% to 8.69%. Those were autocomplete-era effects. The new paper’s 180% is the agent-era version of the same measurement problem: activity moves first, shipped software later.

METR’s follow-up in late 2025 muddied the slowdown result rather than reversing the NBER fade. Among the original developers who came back, the point estimate flipped to an 18% speedup, with a confidence interval from 38% faster to 9% slower. New recruits were about 4% faster, again with a wide band. Many people now refuse to work a study task without AI, so the trial is missing the heaviest users. The observational GitHub panel, which does not ask anyone to put the tools down, is the cleaner read on what shipped after they kept them on.

Smaller Teams Still Face the Same Last Mile

Demirer’s advice to firms is operational, not lyrical. Because coding throughput rose, teams can be smaller, with fewer handoffs. The hours that come back should move to merge, release, and upkeep, the stages that still sit with people. Leaders can also point tools at the paperwork around those stages, such as meeting notes and approval queues, if the block is process rather than taste.

“Some bottlenecks are truly technical and mechanical tasks” that people still have to finish because the models are weak at them, Demirer said. “As long as process aren’t defined by technical or institutional knowledge, AI tools can resolve them in a very particular way.” The grammar in that second line is his. The point is not. If the remaining work is judgment, taste, or liability, a faster coder does not clear it.

That is the hidden labor the commit charts hide. Reviewers, release managers, store-trust teams, and the users who have to find the new binary are the ones absorbing the 180%. IT groups are already treating employee-built vibe apps as a fresh attack surface, because a business unit can now stand up something on production data with no review trail. Security is another downstream queue. It does not show up in GitHub commit counts either.

The paper is also clear about what it cannot see. Quality is inferred from ratings and downloads, not from reading the code. Closed-source and internal software are out of the panel. Software is the most mature generative-AI market, so the 30% release figure may not travel to writing, law, or support. Those caveats cut both ways. Internal tools could be where the coding boom becomes real shipped value. They could also be where unread agent drafts pile up in private repos instead of public stores.

A 30 Percent Release Gain Is Still Rare

Musolff will not throw the result out. “If the world froze at today’s level of AI capabilities, these results would be a bit of a cold shower,” he said. He also said a 30% rise in software releases is still the kind of return few other tools can sell. The authors had studied earlier Copilot builds; he called the gap between those and the agent stack night and day.

In software, the binding constraint appears to be shifting from writing code to reviewing, integrating, and ultimately distributing it.

Mert Demirer, Leon Musolff, and Liyuan Yang, NBER Working Paper 35275

If later models write code that needs less review, or if they take over more of merge and release, the 180-to-30 fade can narrow. That is a claim about the next generation, not a reading of this one. This generation filled GitHub and then filled the stores. On the four marketplaces in the paper, usage in the first three months after launch still has not risen with that supply.