Every few months, a founder posts that they shipped an app over a weekend for forty dollars in API credits. A few days later, an engineer replies with a screenshot of a four-thousand-dollar bill from the same kind of tool. Both are telling the truth, which is exactly why “is AI cheap enough to replace developers?” is harder than it looks.
Here is the short version. AI is cheap enough to change how many developers you need and what they spend their days doing. It is not yet cheap enough, or reliable enough, to replace the people who decide what to build, check that it works, and answer when it breaks. The reason is simple: the price of the tool and the price of a finished, trustworthy outcome are two very different numbers.
The sticker-price case is real
Start with what a developer costs. The US Bureau of Labor Statistics put the median software developer wage at $135,980 in May 2025. A common rule of thumb adds roughly 30% for benefits and payroll taxes, and that’s before recruiting costs and the weeks a new hire takes to get productive. A US developer therefore costs somewhere around $175,000 a year, or about $14,700 a month.
Now the AI side. Individual plans start around $20 a month, and for most solo developers the realistic range is $20 to $100. For enterprise teams, figures attributed to Anthropic and relayed in Morph’s 2026 cost analysis put typical Claude Code spend at about $13 per developer per active day, or $150 to $250 a month. One 2026 comparison pegs the effective cost of AI coding at $1 to $8 per hour, against $40 to $250 an hour for human developers.
On those numbers, AI looks 10 to 50 times cheaper. If that were the whole story, this article would be one sentence long. It isn’t, for three reasons: the bill is no longer flat, the output is harder to measure than it looks, and a hidden cost has moved from writing code to checking it.
Reason one: the bill stopped being flat
Coding assistants used to be seat licenses, a fixed fee per developer per month. The major vendors have since moved toward consumption pricing, where you pay for the tokens your agents burn. Gartner analyst Nitish Tyagi told The Register in June that bills are jumping from $20 or $100 a month to $2,000–$5,000 per developer, with extreme cases reaching $20,000. Gartner projects that by 2028, AI coding costs will overtake the average developer’s salary.
That forecast needs context. Gartner is not claiming it will happen everywhere, since US salaries sit well above those in India. But token prices don’t vary by country, so a heavy user in India may already cost as much as an engineer with four to six years of experience.
Why are bills rising when per-token prices keep falling? Because usage is growing faster than prices are dropping. Gartner estimates that agentic workloads consume 5 to 30 times more tokens per task than a chatbot exchange, since an agent reads files, plans, edits, runs tests, fails and tries again. It’s a Jevons-style effect: when a unit of work gets cheaper, people do vastly more of it.
There’s also a behavioral trap. Amazon shut down an internal leaderboard that ranked employees by token usage after people launched junk tasks to climb it, a pattern now called “tokenmaxxing.” Gartner’s own finding is that burning more tokens doesn’t reliably produce more productivity. Measure activity, and you get activity.
To keep this in proportion: even at the heavy end of $5,000 a month, AI costs about $60,000 a year, well under a US developer’s loaded cost. But it’s a variable cost, hard to forecast, set by vendors, and capable of ballooning. A salary doesn’t do that.
Reason two: what are you actually getting for the money?
Cost is only half the ratio. The other half is output, and this is where the evidence gets messy.
The most rigorous study so far is a 2025 randomized trial by METR, a nonprofit that evaluates AI systems. Sixteen experienced open-source developers worked through 246 real tasks on codebases they knew well. With AI tools they took 19% longer, yet afterward they still believed AI had made them about 20% faster.
METR’s 2026 follow-up complicated the picture. Estimated speedups appeared (around 18% for returning developers and 4% for new ones), but the confidence intervals include zero, and METR itself called the results unreliable. The reason is telling: 30% to 50% of participants admitted they were holding back tasks they didn’t want to do without AI, which skews the sample toward work where AI helps least. METR says speedups now seem likely, but it can’t yet say how large.
Self-reports are more generous. In a METR survey of 349 technical workers, the median person claimed a 3x speedup, but only a 1.4x to 2x increase in the value of their work, and METR warned those answers may not be grounded in reality. Stack Overflow’s 2026 survey, with more than 30,000 responses, found that 65.9% of developers use coding assistants or agents. Yet only 6.6% would trust AI with an important decision, and roughly half trust its output only when they can verify it themselves.
The honest reading: AI helps, probably a lot on some tasks, but nobody has a trustworthy single number, and people are bad at judging their own speedup. Dividing a precise cost by a fuzzy benefit gives you a fuzzy answer.
Reason three: the verification tax
Writing code got cheap. Knowing the code is right did not.
A December 2025 analysis of 470 open-source pull requests by CodeRabbit found AI-co-authored changes carried about 1.7 times more issues than human-only ones, including 75% more logic and correctness problems and 1.5 to 2 times more security findings. Treat that with care: CodeRabbit sells code-review tools and the sample is small. But it matches what other reports describe. An Opsera analysis found AI cut the time to open a pull request by roughly half, yet those pull requests then sat in review queues about 4.6 times longer.
That is the pattern to watch: the bottleneck moves from typing to reviewing. If agents produce code faster than your senior engineers can verify it, you haven’t removed a cost; you’ve relocated it to your most expensive people. Amazon’s own internal guidance on AI in engineering, as described this week by columnist Corey Quinn, advises looking at the teams with the highest token consumption and checking whether their delivery actually rose in proportion. Few companies are doing that check.
What the job market says
If AI were already cheaper than developers across the board, you would expect developer employment to be collapsing. It isn’t, but something is shifting.
The BLS still projects 10% growth in software developer jobs from 2025 to 2035. Indeed’s data shows US software postings about 27.5% below their pre-pandemic level but recovering, and senior roles accounted for roughly 71% of the increase between May 2025 and May 2026. Meanwhile entry-level roles made up just 4.5% of software postings in early 2026, and Stanford’s 2026 AI Index reports that employment of software developers aged 22 to 25 is down nearly 20% from its 2024 peak. Tech layoffs have already passed 133,000 in 2026, more than all of 2025, according to Layoffs.fyi.
Causation is contested. LinkedIn’s economists argued earlier this year that interest rates explain most of the swings in software hiring. But the shape of the shift is hard to miss: AI isn’t replacing developers wholesale, it’s thinning out the first rung of the ladder. GitHub says about one in three pull requests now involves an AI agent, and companies seem to be hiring fewer people to do the work agents handle and more experienced people to direct and check it.
The wildcard: the capability curve
Everything above describes today. The reason the debate is so heated is that the trend line is steep.
METR measures how long a task, in human expert hours, an AI can finish at 50% reliability. Its January 2026 benchmark put Claude Opus 4.6 at roughly 14.5 hours. The UK’s AI Security Institute, citing METR, estimates software task length has been doubling about every 4.2 months since late 2024.
Two caveats matter. First, 50% reliability means the agent fails half the time, and nobody keeps an employee who botches half their assignments. Reliability, not task length, is the gate to real replacement. Second, these benchmarks are mostly self-contained, well-specified tasks, not the ambiguity, politics and legacy code of real software work. But if reliability climbs toward the high nineties, the verification tax shrinks and the math changes quickly. Past predictions about what AI can’t do have aged badly, so humility is warranted.
A better way to do the math
Stop comparing salary to subscription. Compare the cost of a shipped, verified, maintained outcome:
(AI spend + human review and rework + incident cost) ÷ changes that actually reach production and stay there
Run that for a human, for AI alone, and for a human using AI. Most teams will find the third column wins.
| What the sticker price counts | What it leaves out | |
|---|---|---|
| Human developer | Salary and benefits (about $175,000 a year in the US) | Hiring, ramp-up, management time |
| AI agent | Subscription or tokens ($20 to $5,000+ a month) | Review, rework, security checks, retries, and someone accountable |
A few practical moves follow from this:
Track cost per merged change, not tokens. Tokens measure effort; merged, working changes measure results.
Route work by difficulty. Gartner recommends sending simple, high-volume tasks to smaller models and reserving frontier models for hard problems, and improving the context you give agents rather than just feeding them more.
Set spending caps. A runaway agent session can cost thousands of dollars in days.
Invest in automated tests and security scanning. They make verification cheaper, which is where the real savings hide.
Don’t burn the pipeline. If you stop hiring juniors entirely, you’ll be short of seniors in five years.
The verdict
Is AI cheap enough to replace developers? For narrow, well-specified work, in the hands of a skilled person who can check the result, it is already cheap enough to replace a lot of hours of developer time. It is not demonstrably cheap enough to replace developers: the people who define the problem, judge the trade-offs, verify the output and own the consequences. Those parts have not gotten cheaper, and as AI produces more code, they become the scarcest part of the process.
The more useful forecast is a reshaping, not a replacement: smaller teams, a higher bar to enter, more weight on judgment, and a budget line for tokens that finance teams will learn to scrutinize like cloud spend. Ask the question again in a year. Both the price and the reliability will have moved, and those are the two numbers that decide it.
Sources
US Bureau of Labor Statistics, Software Developers (median wage, 2025–35 outlook)
The Register, Corey Quinn on Amazon’s AI memo and tokenmaxxing (October 2026)
AI Cost Estimator, AI coding cost per hour vs developer rates
CodeRabbit, State of AI vs Human Code Generation (Business Wire)
Opsera and Sourcery 2026 reports, as summarized by a Codex CLI developer blog
Tech Hiring Statistics 2026 (Indeed Hiring Lab and Layoffs.fyi data)
UK AI Security Institute, how fast is autonomous AI cyber capability advancing (citing METR)
Published via ZyVOP — Write once in Markdown, auto-backup to GitHub, and syndicate to Dev.to, Medium & Hashnode in 1 click.
About the author
- One platform to write, cross-post to Dev.to, Medium, WordPress, Hashnode, and Bluesky, and protect your SEO with canonical links. Zero paywalls and full content ownership.
Comments
Join the discussionNo comments yet. Be the first to comment.
