Does Coding with AI Really Save Time and Money?
ai #agents #programming #productivity
Hard research meets token-fueled coffee chat humor
How much has AI truly reshaped our daily workflow? Are coding agents actually turning us into speed demons? Instead of relying on gut feelings and hot takes, I wanted to see what actual scientific studies have to say.
Like my last piece, this post was sparked by the upcoming AI Coding Summit in Berlin (yes, I’ll drop my speaker badge link at the bottom again—it might even snag you a discount). Despite swearing I wouldn't overcommit again, I somehow ended up on three panels. 😅 One of them tackles almost this exact question.
Since I wanted real evidence rather than vibes for that panel, I went down a research rabbit hole. Every paper I mention is linked at the end.
Plus, today has been a bit of a rough one and I’d honestly rather be anywhere else, so I decided to lift my spirits the way I always do: by writing something for you!
Are we actually shipping faster?
Let’s tackle speed first. If you cornered me on the street and demanded a quick answer, I’d say yes—we are definitely cranking out code faster than before.
The catch is that faster typing hasn't magically doubled our output of finished software. We still have to deal with code reviews, architecture discussions, analysis, and everything else wrapped around writing code. The day-to-day feels a bit smoother and maybe slightly less stressful, sure. Still, the end-to-end delivery pipeline hasn't exactly shifted into hyperdrive.
Then again, maybe it’s just my corner of the industry. Is everyone else shipping features at warp speed?
Turns out, not really.
A study by economists at MIT and Wharton (revised in September 2026) backs up that hunch. After examining data from over 500,000 GitHub developers, they estimated the following cumulative shifts linked to autonomous coding agents:
| Metric | Estimated Change |
|---|---|
| Coding activity (commits) | +240% |
| Software releases | +30% |
| Total app usage across studied marketplaces | No overall increase |
Keep in mind these are observational estimates, not controlled lab results. Even so, the gap between skyrocketing commits and modest release growth says a lot.
In short, we’re pushing way more code and releasing slightly more often. Personally, it mostly feels like it takes a bit of the edge off the workload rather than sparking a total revolution. Greenfield projects might be a different story, of course. On the large, mature systems I work on, I haven't witnessed anything resembling a 10x leap.
Maybe that changes once someone cracks the code on truly self-healing autonomous agents. Right now, though, things sometimes feel more tangled—especially with the recent wave of AI security concerns making everyone (myself definitely included) a bit paranoid. 😅
Are individual developers actually faster?
Fine, commit counts are up—but are individual engineers genuinely completing work quicker? Again, my gut says yes, just not by an earth-shattering margin.
The research paints a mixed picture:
| Study | Result | What Was Measured |
|---|---|---|
| Cui et al., Management Science, 2026 | +26.08% | Number of completed tasks |
| METR, July 2025 | 19% slower | Time to complete tasks |
| Science, January 2026 | +3.6% | Estimated increase in quarterly GitHub coding activity |
| Borg et al., June 2026 | 30.7% less time | Median drop in implementation time (observational phase) |
Once more, this tracks pretty closely with real-world intuition. I’ll spare you ten paragraphs dissecting methodologies—check the links below if you want the weeds. Each paper tracked different tools, developers, and tasks, so those percentages aren't apples-to-apples.
Still, most of them point to genuine gains. None of us just mutated into 10x engineers overnight xD, yet the productivity bumps are real.
METR is the glaring exception. Published in 2025, it practically qualifies as ancient history on the AI timeline, yet it uncovered a fascinating psychological quirk.
Going into the study, developers predicted AI would speed them up by around 24%. In reality, it slowed them down by 19%. The funniest part? Even after finishing the experiment, they still walked away convinced AI had made them faster.
It’s a humbling reminder of how easily we fool ourselves without hard measurements—a truth that applies way beyond software. (For years, I swore I barely ate anything. 😂)
To METR’s credit, they released an update in February 2026 noting that newer models are probably more useful, though participant self-selection made it impossible to calculate a reliable speedup figure in their follow-up.
Okay, but is development actually cheaper?
If you shook me awake at midnight and asked whether AI has lowered the cost of building software, I’d tell you no—at least not yet.
Both the 2025 DORA report and its March 2026 follow-up highlight a pattern many of us have noticed firsthand: a chunk of the time we save generating code immediately gets eaten up by auditing, verifying, and double-checking it. Implementation savings can evaporate fast on the road to production. DORA maps out how that effort shifts rather than disappears, leaving the true bottom-line impact unresolved.
And yes, I know all about multi-agent pipelines where one agent codes, a second reviews, a third tests, and everything ships straight to production. That might fly in low-stakes environments. Trying that in mission-critical software—say, financial systems? Best of luck with that! 😂🤞
Are token bills wiping out the savings?
One final question: are we burning all our time-savings money on API tokens?
Honestly, the research caught me off guard here. I would have bet yes, especially if teams throw frontier models at every trivial prompt.
The reality looks a bit more nuanced. Two studies tackled the economics directly—one by Potluri and colleagues in September 2026, and another by Steffen Herbold in October 2026.
Both point to the same conclusion: agent capability matters more than token pricing. In Potluri’s analysis, stronger models that actually solved tasks drove substantial estimated savings, while weaker models frequently ended up more expensive across many categories. Cheap tokens add up fast when a model keeps botching the job and forcing retries.
Herbold took a different angle, running 432 simulations of hybrid human-AI teams that factored in salaries, token spend, oversight, bugs, and technical debt. Teams paired with highly capable agents typically beat human-only teams on cost, whereas weaker agents produced mixed results. Capability beat low token prices because higher-quality output slashed the time spent reviewing and fixing broken code.
Of course, Herbold’s numbers come from simulations and Potluri’s rely on benchmarks and economic modeling. They offer useful clues about when AI pays for itself, even though the full lifecycle cost of commercial software remains a much thornier equation.
The bottom line
Taken together, the evidence suggests AI can make coding both faster and cheaper. Multiple studies report measurable speedups on discrete tasks, and economic models indicate that strong agents can generate net savings even after factoring in token bills.
What’s still missing is solid proof that the total cost of building, shipping, securing, and maintaining software has dropped across the board.
For now, the science mostly mirrors what we feel in the trenches: Skynet may be sharpening its knives, but it hasn't taken over the world just yet. 😅
How does this stack up against your day-to-day? Is AI genuinely making your team faster and leaner, or are you seeing a completely different reality at work?
Drop your thoughts—I’d love to hear them!
If you enjoyed this read, feel free to follow me on LinkedIn.
And if you want free resources from the AI Coding Summit, you can sign up through my speaker badge.
References
- https://www.nber.org/papers/w35275
- https://doi.org/10.1287/mnsc.2025.00535
- https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
- https://metr.org/blog/2026-02-24-uplift-update/
- https://doi.org/10.1126/science.adz9311
- https://doi.org/10.1007/s10664-026-10889-1
- https://dora.dev/research/2025/dora-report/
- https://dora.dev/insights/balancing-ai-tensions/
- https://doi.org/10.13021/jssr2026.5687
- https://arxiv.org/abs/2610.20394
Top comments (0)