Technical Context
I focused not on abstract complaints, but on a very down-to-earth case: you ask Claude Code in the morning to commit the previous evening's changes, and 6% of your five-hour quota vanishes from a fresh start. On Max 5x, this isn't a minor discrepancy; it's a direct hit to the workflow.
Based on the Max 5x plan, we're talking about roughly 88,000 tokens in a 5-hour rolling window. Six percent of that is about 5,300 tokens. For an operation like "gather the diff, create a proper message, and commit," that number seems wildly inflated.
I dug into how users and observers of the Anthropic ecosystem explain this. A pattern emerges: a cold start pulls extra context into the session, prompt caching is sometimes unstable, and automatic retries and internal service steps discreetly bloat consumption.
So, the problem isn't a single bad request. It seems Claude Code has high baseline overhead at the start of a session, which is especially noticeable on simple tasks. When a tool spends thousands of tokens before it has done any real useful work, the economics start to creak.
A separate nuance is that the quota is a five-hour window, not a monthly one. You can hit the ceiling in the first half of the day, and then you're not reallocating a budget—you're just sitting and waiting. For development, this is more frustrating than a typical pay-per-use model because it blocks the actual workflow.
Anthropic has already acknowledged that some users are hitting limits faster than expected. But from an engineer's perspective, what matters isn't the acknowledgment itself, but whether the behavior has been fixed in practice. Current feedback suggests that for some teams, the answer is still no, it hasn't been fixed.
Impact on Business and Automation
What gets me here isn't the daily inconvenience, but the architectural implication. If a simple operation like a commit unpredictably burns a significant chunk of the window, I can't confidently design a reliable AI integration into the dev process, CI, or internal agent scripts based on it.
In demos, everything looks great. In production, you suddenly find that the agent meant to save time has itself become a scarce resource.
This is particularly damaging for those looking to build AI automation around routine engineering tasks: generating commit messages, parsing diffs, code reviews, changelogs, and triaging issues. When the cost of a single step fluctuates wildly due to a cold start or a broken cache, budget predictability vanishes. And without it, AI implementation quickly turns into an expensive experiment.
Who wins in this situation? Tools with a clearer economic model: pay-per-token without strict windows, stable caching, transparent limits, and proper telemetry. It's no surprise that people are already looking at Codex and other alternatives where it's at least easier to understand what you're paying for.
Who loses? Teams that have tied critical parts of their process to a single agent-based tool without a backup plan. I've seen this before: first, everyone celebrates the speed, then they hit the limits mid-sprint, and it's back to manual firefighting mode.
This is precisely why at Nahornyi AI Lab, we generally don't build AI solution architectures around a single vendor and a single pricing model. For businesses, I almost always incorporate fallback routes, separate pipelines for expensive and cheap tasks, context caching, and strict control over unit economics. Otherwise, any AI architecture crumbles at the first spike in limits.
My conclusion is simple: Claude Code can still be a useful tool, but if basic operations on Max 5x start eating up 6% of the window, it's no longer a minor bug but a signal to reassess your stack. For personal use, it's an annoyance. For a business, it's a risk that must be calculated in advance.
This analysis was done by me, Vadym Nahornyi of Nahornyi AI Lab. I don't just repeat press releases; I gather and test these things in real-world AI automation and AI solution development scenarios for teams.
If you'd like, I can help analyze your case: where you're currently burning tokens, how to restructure your AI implementation, and what can be used to back up critical processes. Contact me, and let's look at your project together.