Most professionals running AI seriously are paying for three to five subscriptions. OpenAI, Anthropic, Cursor, Perplexity, maybe a few API keys on top. The monthly total adds up fast and almost nobody has a clear picture of what they are actually using versus what they are paying for.
But the subscription cost is not the real problem. The real problem is the waste happening inside each session that nobody talks about.
Token waste is invisible until you measure it. And most people never measure it.
Here are the three biggest drains and exactly how to fix them.
Drain 01 — The Follow-Up Tax
Every time you send a correction message after a bad output, you are paying twice. The first prompt consumed tokens. The follow-up message adds to the context window and consumes more. Then the corrected response consumes more again. Three rounds of token spend for what should have been one. The correction also adds noise to the context window, which compounds into worse outputs further down the session.
Drain 02 — The Single-Task Habit
Every new prompt you send carries a context loading overhead. Three separate prompts for three related tasks cost significantly more than one well-structured prompt that batches all three. Most people do not do this because they have never seen the cost difference. When you see the numbers in real time, the habit changes immediately.
Drain 03 — The Subscription Blind Spot
You are either hitting rate limits and losing time, or you are underutilising a subscription and wasting money. Very few people know which is true for each tool they pay for. A Claude Web subscription with a 5-hour usage window you consistently cap out on is a different problem than a Cursor API you are using at 30% of its monthly limit. Without visibility, you cannot fix either.
01
Edit, do not follow up.
When a response misses the mark, go back and edit the original prompt rather than sending a correction. Most AI interfaces support this. It resets the context from that point cleanly, costs fewer tokens, and keeps the context window free of noise. One prompt, one response. That is the target.
02
Batch related tasks into one prompt.
If you need a summary, three subject line options, and a shortened version of the same piece of content, put all three requests in one prompt. The model handles context loading once instead of three times. The output is also more consistent because all three tasks share the same context. Faster, cheaper, and better quality.
03
Schedule heavy work outside peak hours.
Anthropic and OpenAI both experience higher load between 5 AM and 11 AM Pacific time. During these windows, rate limits are hit faster and response quality can dip. If you are running intensive AI workflows, shifting them outside this window costs nothing and gets you more consistent output at the same spend.
Tokens 4 Breakfast - Visit Website
Use this to run a three-minute AI spend audit. Give it an honest picture of how you currently use AI and it will tell you exactly where the waste is:
I want to audit my AI usage for cost and efficiency.
Here is my current setup:
- Tools I pay for: [list your subscriptions and approximate monthly cost]
- How I typically use each: [describe your main use cases per tool]
- My usual session pattern: [e.g. "long back-and-forth conversations" / "lots of quick one-off prompts" / "mix of both"]
For each tool, tell me:
1. Whether I am likely over-paying or under-utilising based on my usage pattern
2. The single highest-impact change I could make to reduce waste
3. Whether a cheaper or different plan would serve my actual usage better
Be specific. I want numbers and concrete actions, not general advice.
The output will not be perfect without real usage data, but it surfaces the questions most people have never asked about their own AI stack. Run it once and you will know exactly which subscription to cut, which to upgrade, and which habit to change first.
If you want to go deeper on getting more value from the AI tools you already pay for, the free Prompting Hub at tminusai.com covers 19 laws of prompting and 32 cheat codes that reduce wasted output at the session level. Less re-prompting means fewer tokens consumed. Better inputs produce better outputs on the first pass.
The Power Guide Pro covers the full system: session architecture, Projects setup, and a weekly workflow that compounds. One purchase, permanent reference. Euro 59.99.
And if you are on a Mac and running multiple AI tools, Tokens 4 Breakfast is the fastest way to see exactly what your stack is costing you in real time.
More next week.
Kapish
Enterprise AI Architect · T-Minus AI