How I Keep My AI Coding Agent Costs Down as a Student

I graduated from undergrad the year that ChatGPT was first released to the public, and I'm now applying to Master's programs for the Fall. As we inch away from the educational realm I knew, an era ruled by StackOverflow and Chegg, I can't help but think how different my learning path would have been had I had access to these tools back then. Depending on your self-control, AI can either give you a competitive edge, even over others using LLMs for everyday tasks, or it can sneakily lead your learning to a plateau. But, with great power, comes great responsibility. And no, I'm not just talking about the learning cost of using AI maliciously in an academic context (i.e. cheating on exams or labs), but also in financial cost. Every prompt, response, and tool call burns tokens, and tokens cost money. And given the control Chipotle had over my bank account in college, I was sticking to free tiers wherever I could. I assume you are too. For the money you can spare, however, AI coding agents can build complex apps, explain concepts, and create study resources to help you reinforce your learning. Well-scoped prompts, frugal model choice and appropriate MCP server usage can use a tenth of the tokens of a rambling prompt, maximizing value per request and stretching out that usage window each month. After applying to some Master's programs, and coding extensively in conjunction with editors like Kiro and Claude Code, I've settled on a few habits that postpones that dreaded "Try again tomorrow at 5:30pm" message in my chat window. Here's three of those habits. 1. Prompt Engineering (Yes, it's a real skill) I'll be honest... when LLMs first started gaining traction and people started talking about "Prompt Engineering", it felt like a snobby title. I thought "you know what chatgpt is, and you know how to write... so what?". Comedians were going on Jimmy Kimmel saying they asked ChatGPT what it thought, my roommate was asking ChatGPT how to make a comeback to my joke... it just seemed like people considered them more technical for knowing how to ask a genie a question. But this is 2026. Models are smarter, MCP reduces model guesswork, and the software industry is innovating faster than ever. As such, there is MUCH to be gained in reducing your token usage, and Prompt Engineering encompasses that first critical skill; accurately and concisely structuring your request. With that in mind, the most expensive prompt is one where the agent has to guess. So, when prompting, you should... Show up with a plan, not a blank page. I don't ask the agent to design my whole system if I already know roughly what I want. If my assignment calls for a REST API with a database and a couple of endpoints, I say that directly. I don't make the agent spend tokens rediscovering decisions I've already made in my head. Say what you're sure about, and flag what you're not. If your class requires Python and Flask, say so up front. If you genuinely don't know whether to use SQLite or Postgres for a small project, say that's the part you want the agent's input on. This tells the agent where to just execute and where to actually reason, and reasoning costs more tokens. "Build me an app" forces the agent to guess at dozens of small decisions and explain each one. "Build me a Flask API with these three routes, and suggest the best way to handle rate limiting" puts the expensive thinking exactly where you need it and nowhere else. When I'm unsure which tool is best, I'll just do some research with Gemini to try and make a decision before asking my coding agent. Compress your prompt with a free model first. Before sending a long, complicated prompt to your paid agent, paste it into a free-tier model (ChatGPT's free tier, Gemini, whatever you have access to) and ask it to tighten the prompt without losing any meaning. You're paying by the input token, so a shorter prompt with the same information is money back in your pocket. 2. Match the Model to the Job Not every task needs the most powerful model available. Picking the right model for each task is the simplest way to optimize your inference spending. Look up how the models compare on your current task. Benchmarks like SWE-bench and HumanEval show which models are strongest at writing and debugging code versus which ones are better at prose. Know the strengths of whatever models you have access to. Check what each model actually costs you. In Kiro (an AI coding IDE), for example, you can see how much credit each model uses relative to "Auto" mode, which picks a model automatically based on how complex your prompt looks. Auto is convenient when you're not sure, but knowing the relative costs lets you choose deliberately when you already know the task is simple or hard. Match the model to the job: A frontier model (like Claude Opus 5, GPT Sol) is worth the extra cost for multi-file projects, tricky debugging, or anything where the agent needs to hold a lot of context at once. Yes, it costs more per token. But a wrong implementation that takes three rounds of fixing costs more overall than getting it right the first time. A lighter model (like Haiku) is plenty for writing docs, generating boilerplate, drafting a README, or formatting code. These tasks don't need deep reasoning, and a cheaper model handles them just as well. The rule of thumb: if the task requires the agent to think, use a strong model. If the task just requires the agent to type, use a cheap one. 3. MCP Tools: Let the Agent Check Instead of Guess Model Context Protocol (MCP) servers are probably the most overlooked way to cut costs. An MCP tool gives your agent direct access to real information (documentation, live APIs, code search) instead of making it guess based on whatever it remembers from training. Without a tool, an agent has two options when it needs a fact it doesn't have memorized: guess and risk being wrong, or ask you and burn your time and more prompt tokens. An MCP tool call is cheap: a quick lookup that returns the exact right answer. Compare that to the alternative, where the agent writes a plausible-looking but wrong API call, you catch it, you explain the fix, and the agent tries again. That round trip alone can cost four times as many tokens as just looking the answer up in the first place. Two that I use regularly: AWS Documentation MCP gives the agent direct search access to current AWS docs, so it's not relying on training data that might be stale for newer services. Strands MCP helps your agent build other agents. If you're working on a multi-agent project, this saves you from having to paste boilerplate framework code into context repeatedly. A single documentation lookup might cost a few hundred tokens. The alternative (wrong answer, correction, retry) easily runs into the thousands. Over a semester of daily use, that adds up fast. Keeping an Eye on What You're Spending The three habits above reduce cost. This last one is about actually seeing it, because you can't fix what you can't measure. Kiro's dashboard updates every few minutes and shows your credit usage against your monthly pool, plus a per-interaction cost right in the notification bar. That's an immediate feedback loop: send a prompt, see what it cost, adjust next time. For more detail, there's a community tool called kiro-usage (on PyPI) that gives you a live terminal dashboard breaking usage into new context, reused context, and output tokens. It keeps a history even across session restarts. Install it with: uv tool install kiro-usage && kiro-usage install Then run kiro-usage in a side terminal while you work. Once you've watched your usage for a week or two, predicting cost gets easy. Kiro's Auto mode is the baseline (1x), and other models are published as multipliers against it, so if a task usually costs half a credit in Auto, you can estimate what switching models will cost before you commit. You'll also start noticing patterns: a quick bug fix costs one thing, building a whole feature costs several times that, and pasting a 500-line file into context costs more than pointing the agent at the 30 lines it actually needs (roughly one token per four characters, as a rule of thumb). The Point of All This Overall, the name of the game is to reduce your agent's guesswork. Every design decision you make concretely, every MCP tool you set up, and every unnecessary word you eliminate from your prompt all reduce the amount of inference your agent has to perform (and charge you for) to complete your task. You end up with the same quality of work for a fraction of the credits, which matters a lot more when you're working on a student budget as opposed to an enterprise. Employing these skills in your project will position you well for a software industry where AI is critical infrastructure. Thanks for tuning in.