A new analysis argues that using large language models for coding is still far more expensive than most users realize, even when subscription plans make the experience appear affordable.
The piece, based on months of hands-on experimentation with Claude Code, says the economics of LLM-assisted software development are being obscured by subsidized consumer pricing and by the gap between visible token charges and the actual compute needed to complete complex tasks.
The author says a $20 monthly plan quickly ran into usage caps, and that buying extra capacity at API rates made the effective cost much higher. After adding roughly $80 in purchased tokens, the conclusion was that a higher-priced subscription, such as a $100 monthly plan, looked like a much better deal than paying incremental API prices whenever the plan limits were reached.
That experience led to a broader cost analysis of LLM use in coding. According to the article, the key issue is not the listed price per token, but the total cost required to finish a task. The author argues that coding workloads often need repeated back-and-forth prompting, tool use, retries, and additional hidden computation, all of which can multiply token consumption.
The analysis estimates that a user on a premium Claude plan who pushes the system to its weekly limits through highly agentic coding could consume tokens worth more than $1,000 at API pricing. The article says this suggests current subscription models are still absorbing much of the underlying cost.
The author distinguishes between simple chatbot use and more demanding applications such as coding and complex reasoning. Simple questions, the article says, can now be so cheap that the marginal cost is nearly negligible. But more involved tasks rely on recursive processes, tool use, and repeated internal attempts to improve outputs, which can consume far more tokens than users see.
The article also says one high-effort task on a frontier model may cost around $75 at API rates, and that a single query can sometimes run to about one million tokens. That, the author argues, makes the economics of serious LLM coding very different from casual chat use.
In the article’s view, vendors may be promoting an image of broad affordability by emphasizing low per-token prices and simple-use cases, while the real cost profile for demanding workloads is much higher. The author says that for coding in particular, where accuracy matters and mistakes can be expensive, the practical cost of getting a working result is what matters most.
The analysis also questions whether recent model improvements will be enough to fix the problem. The author says Anthropic appears to be working on newer versions that reduce token use, but even if those efforts succeed without sacrificing quality, they may also signal that the current growth pattern is nearing its limit.
More broadly, the article frames LLM-assisted coding as a powerful but still immature tool. It says the technology has enabled the creation of a functional application that would otherwise have taken far longer or might not have been built at all, but argues that the economics remain fragile.
The bottom line, according to the author, is that LLM coding is useful and impressive, but not yet cheaply scalable for most serious use cases. For now, the article says, the apparent affordability of these systems depends heavily on vendor subsidies and on users not looking too closely at the cost of finishing the job.