Executives once dreamed of AI slashing engineering payrolls. Fire coders. Feed prompts to bots. Watch output surge. But bills are piling up. One engineer torched $150,000 in a single month on Anthropic’s Claude Code, as detailed in a New York Times investigation from March 2026. Reality bites harder than the hype.
AI coding tools promised efficiency. Developers crank code faster. Companies trim headcount. Yet costs now rival salaries. Nvidia’s Bryan Catanzaro put it bluntly: “The cost of compute is far beyond the costs of the employees,” according to a Futurism report published May 3, 2026. Anthropic quietly revised its pricing estimates. Average spend jumped to $13 per developer per active day, or $150-250 monthly. For 90% of users, it’s under $30 daily—but teams run multiple agents. Expenses stack. A single employee can match a human’s full pay in tokens alone.
Tokenmaxxing swept tech firms. Workers compete on leaderboards at Meta and OpenAI, racing to burn the most AI fuel. Software engineer Max Linder admitted, “I probably spend more than my salary on Claude.” Ege Erdil of Mechanize pegged one full-time agent at 700 million tokens weekly. That’s not productivity. That’s a power grab.
And the code? Mostly trash. GitClear’s analysis of 211 million lines from 2023-2025 showed churn—fresh commits deleted or rewritten within two weeks—spiking from 3.3% to 7.1% post-AI. AI-assisted repos logged 39% more churn, per their January 2026 research on GitClear. Duplicates quadrupled. Heavy AI users produced 4-10x more durable code but 9x more churn. Quantity soars. Quality craters.
Security flaws compound the mess. Veracode’s Spring 2026 update tested over 100 LLMs on 80 tasks. Syntax hit 95% accuracy. Secure code? Just 55%. Nearly half introduced OWASP Top 10 vulnerabilities like SQL injection and cross-site scripting, as outlined in their blog. Java fared worst at 72% failure. Models mimic public repos—bugs and all.
Productivity claims falter under scrutiny. METR’s mid-2025 RCT, revised February 2026, found senior developers 19% slower with tools like Cursor and Claude, despite feeling 20% faster—a 39-point perception gap, via their update. NBER’s February 2026 survey of 6,000 executives revealed 80% saw zero impact on productivity or jobs over three years (NBER paper w34836). The overwhelming majority of companies saw zero revenue growth post-AI, echoing Futurism’s MIT reference.
This is the J-Curve in action. Volume peaks first. The Synthesis captured it May 3, 2026: AI metrics tout time savings—Microsoft’s 55% faster tasks, Stanford’s 75% cycle cuts—but ignore rework. DORA 2025 warns healthy turnover stays under 15% at 30 days. AI blows past, breeding technical debt. Snyk flagged 29.1% of AI Python code with security issues.
Teams feel the strain. Juniors get replaced, seniors debug “workslop.” A fintech squad doubled output, cut headcount 30%, then watched incidents rise 28% and resolution double, per LinkedIn analyses echoing GitClear. Amazon lost 6.3 million orders to one AI deployment flop. CVEs from AI code leaped from 6 to 35 in months.
Offshore shifts accelerate. CFOs eye spreadsheets: $340k U.S. senior vs. $31k Bangalore peer with Claude—91% savings, identical velocity. Goldman’s models flag 2.7 million U.S. jobs for “optimization.” But churn follows. Bugs cost 100x more in production without context.
Costs climb further. Anthropic doubled revenue forecasts on agentic tools, per NYT. OpenAI’s Codex quintupled tokens. Google hit 1.3 quadrillion monthly. Yet sustainability wanes. “It doesn’t seem sustainable,” said an anonymous OpenAI staffer.
Some wins emerge. Repetitive tasks shine—BCG’s 64% daily user boost. But complex work? AI hallucinates dependencies, skips docs. Developers spend 63% more time fixing its output. PR queues clog. Burnout brews.
Economics sour. Initial savings evaporate in fixes, debt, bills. Futurism nailed it: deploying AI coders across firms looks questionable. Executives chase leaderboards. Code rots. Investors watch power grids strain. The churn machine grinds on.
But cracks show. Firms track real metrics now—churn rates, debt accumulation. Top performers self-select into AI, masking averages. Juniors vanish from pipelines; no one learns to review bot slop. Klarna rehired after AI cuts backfired.
Path forward? Measure outcomes, not tokens. Pair AI with rigorous review. Target drudgery, not architecture. Otherwise, it’s expensive illusion. Compute outpaces coders. Productivity? Still waiting.
WebProNews is an iEntry Publication