How I Burned Through My $200/Month AI Budget in a Week — And the 9 Disciplines That Stopped It
Some links below are affiliate links — I may earn a small commission, at no extra cost to you, if you sign up for a tool I recommend. More on how I choose what to recommend.
Prices spot-checked on May 23, 2026. AI tools change pricing often — always verify on the vendor's site before signing up. See live pricing.
I have a $200/month AI budget at work for Microsoft Copilot. For three weeks running, it was gone by Friday. Not “running low.” Gone. The platform throttling me, my IDE telling me I’d hit my limit, my workflow ground to a halt.
The first week I assumed it was a bad week. The second week I assumed Copilot’s billing was wrong. By the third week — when I’d noticed the same thing creeping into my personal Claude Pro account at home — I had to admit it wasn’t the tool. It was me. I was using AI the way someone leaves the lights on in every room of the house: not maliciously, not consciously, just out of bad habits I’d never had a reason to examine.
The fix wasn’t using AI less. It was using AI more deliberately. The nine disciplines below are what I figured out. Some of them feel obvious in retrospect; none of them were obvious before I needed them. Together they cut my bill by something like half, with no drop in the actual work I was getting done.
If your AI bill has been creeping up, or you’ve been wondering why your “$20/month is plenty” feels like it’s getting tighter — this guide is for you.
A 60-second crash course in tokens
Skip this section if you already know what a token is.
AI tools bill in tokens. A token is a chunk of text — roughly three-quarters of an English word. The word “hello” is one token; “internationalization” is several. Both your input and the AI’s output count. When you paste a 10-page document and ask a question, you might be sending 3,000 input tokens; if the AI replies with a 500-word answer, that’s another ~700 output tokens. Both add up.
On consumer subscription plans (Copilot, ChatGPT Plus, Claude Pro), you don’t pay per token directly — you pay a flat fee with usage caps. But the caps are token caps, even if the tool tells you “you’ve sent too many messages.” Everything below is about not wasting tokens.
(For a deeper take, see our companion piece on whether you need to pay for AI.)
The 9 disciplines
Compress your context before pasting
The single biggest waste I had: pasting a 30-page PDF when 3 pages would do. The AI doesn't care about the rest of the document; you do. Before you paste, ask yourself: what's the smallest slice of this that contains the answer? Open the doc, copy only the relevant pages or paragraphs, paste that. If you genuinely need the whole document, ask AI to summarize it first, save the summary, and work from that going forward — instead of re-pasting the original every time. A typical context-compression habit saves 40–70% of input tokens on document-heavy work.
Paste text, not screenshots
This one is the most embarrassing waste in my old workflow. A screenshot of an error message, a small table, or a formula can cost the AI 1,000–2,000 tokens to process. The same content as text? About 30. Same applies to tables (paste CSV or tab-separated text, not a screenshot of the spreadsheet), code (paste the code, not a screenshot of the IDE), formulas (type them out), and even long PDFs (export to text first when you can).
The discipline is small: before you screenshot, ask whether you could just copy the text instead. Almost always you can — and you've cut the input cost by 10–50× for that prompt.
Kill long-running chats. Start fresh more often.
This is the discipline almost nobody knows about — and it's the biggest invisible expense. Every turn in a chat re-sends all the previous context to the model. A 40-message chat about everything you've done this week sends the entire 40-message history to the AI on message #41, then again on #42, then #43. The cost per turn grows linearly with chat length, and most of those tokens are stuff you've moved past.
Fix: when you're switching topics, start a fresh chat. When the current chat has drifted, summarize the relevant takeaways, copy them into a fresh chat, continue from there. The discipline feels weird at first (you lose the comfort of one big rolling conversation) but it's the single highest-impact change you can make.
Default to the cheapest model that works for the task
Every AI tool now has multiple model tiers — usually a flagship (GPT-5, Claude Opus, Gemini Pro) and one or more cheaper, faster siblings (GPT-mini, Claude Sonnet/Haiku, Gemini Flash). The flagship costs anywhere from 5x to 30x more per token than the smaller model. Most of what you do with AI doesn't need the flagship. Drafting an email, summarizing a meeting, rewriting a paragraph, asking a quick factual question — Sonnet or Haiku handle these as well as Opus and cost a fraction.
Reserve the flagship for what it's actually good at: long reasoning chains, complex code, nuanced analysis, edge-case judgment. Make the switch a conscious choice, not a default.
Cap output length explicitly. Every time.
Without a length cap, AI rambles. It explains, then re-explains, then summarizes its explanation, then offers caveats. Beautiful prose; expensive tokens. Add a length cap to every prompt that doesn't need to be long: "in under 100 words," "in 3 bullet points," "one paragraph." The output is almost always better with the cap — tighter, more useful, easier to read — and you've cut output tokens by 60–80% on most replies.
Fix the prompt, don't regenerate the answer
When AI gives a bad answer, the urge is to hit "regenerate" or say "try again." Don't. That's the same prompt running again — same context, same cost, similar answer. Instead, identify what was wrong with your prompt and rewrite it. "Be more specific about X." "Don't include Y." "Use this format." A fixed prompt produces the right answer on attempt #2 instead of attempt #5. Three saved regenerations is half your hour's token budget.
Use prompt caching where the tool supports it
If you keep sending the same large chunk of context (a coding-style guide, your company's policy doc, a long brief) at the top of many prompts, look up whether your tool supports prompt caching. Both Claude and ChatGPT have it; it stores the cached portion server-side so subsequent prompts that re-use it are billed at a fraction of the original token cost (typically 10% for cached input).
For consumer chat plans you mostly don't have to think about this. For API or developer-tool use (Cursor, Claude Code, the OpenAI API), it can cut a 5-figure monthly bill in half. Worth 30 minutes of reading the docs for the tool you use most.
Never let an agent run unsupervised
AI agents — automated workflows that call the model many times in a loop — are the easiest way to torch a budget. An agent that gets stuck in a loop, or that branches into unexpected sub-tasks, can burn $50 in 30 minutes while you're at lunch. If you use agents (Claude Code, OpenAI's operator, custom workflows), set hard caps: a max number of steps, a max wall-clock time, a max dollar spend. Watch the first few runs of any new agent task end-to-end. Trust comes from observation, not assumption.
Run a 5-minute weekly cost audit
Friday afternoon. Open your AI tool's billing or usage page. Look at this week's number. Compare to last week. If it's up, ask: which of the eight disciplines above did I let slip? Most expensive habits are obvious the moment you look — you'll remember the day you pasted a 60-page contract three times into three separate chats. The audit takes five minutes and pays for itself in the first week you do it.
Tool-specific tips
The nine disciplines work everywhere. A few extras worth knowing about your specific tool.
The mental shift
When to upgrade vs when to discipline
A real moment: sometimes your bill is high because you genuinely need the bigger tier. Not every cost is wasted. Quick framework:
The trap is upgrading first and disciplining never. That’s how monthly AI spend grows from $20 to $200 to $2,000 — not because the work demands it, but because the bill went up so the budget went up.
Did one of these disciplines cut your bill — or is there one I missed? Email help@aiforyourday.com. I update this guide quarterly based on what real readers tell me works.