DEVOURED
Anthropic/OpenAI may be spending more than $1,000 for every $100 you pay them
Heavily subsidized LLM subscriptions are hiding a massive 'brute force' cost explosion in agentic coding tasks.
What: Researcher Renske Wierda analyzed that while simple chat is cheap, agentic coding tasks using 'thinking' models can burn thousands of tokens per turn, making them economically unviable at current API prices without venture-backed subsidies.
Why it matters: The economic reality of AI-assisted development is masked by flat-rate pricing, creating a dangerous dependency on services that may become prohibitively expensive once profit-seeking models take hold post-IPO.
Takeaway: Account for 'thinking' token inflation in your project budget, as Claude 4.7's tokenizer changes and recursive reasoning can increase costs by 35% or more per turn.
Deep dive
- LLM-based coding is currently subsidized; actual costs at API pricing far exceed $100/month subscriptions.
- 'Recursive' or 'thinking' models generate invisible tokens through trial-and-error, significantly inflating usage costs.
- Coding is a 'fault-intolerant' task, requiring more expensive, reliable model efforts than casual chat.
- As codebases grow, complex multi-file changes drive token usage into the millions per quest.
- Scaling models does not inherently lead to intelligence; it leads to exponential compute consumption.
- Real cost per task has exploded, with estimates reaching $65+ per complex code change.
Decoder
- Recursive/Thinking Models: Models that use internal loops, indirection, and self-correction steps (often invisible to the user) to refine answers, consuming significant extra tokens.
- Brute Force Coding: A method where an LLM repeatedly generates, runs, and debugs code until it compiles or passes tests, relying on sheer computational volume rather than reasoning accuracy.
- Fault-Intolerant Tasks: Activities (like coding or medical diagnosis) where small errors have critical, system-wide consequences.
Original article
Full article content is not available for inline reading.
Explore all 68 articles in today's digest
Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi