Written with AI assistance from the source video. The video is the authority.
6:48Better Stack
Log in to leave a comment
No posts yet
When building and operating a web service alone with a monthly budget under 5 million KRW, relying solely on top-tier AI models will cause your API fees to skyrocket in no time. If you choose a coding assistant model based solely on benchmark scores, you'll exceed your spending limit within just a few turns in an actual development environment. By tracking token consumption per task stage and splitting tasks across different models, you can maintain development speed while cutting API expenditures by over 30%.
Web service development is divided into prototyping, core logic writing, refactoring, and debugging. Assuming an average of 50 calls per day and 1,000 API calls per month (based on 20 days), the proportion of input tokens varies significantly across task stages. The debugging phase, where error logs and previous code accumulate in a single chat window, sees input tokens soar up to 45,000 per call, and backend logic writing also demands high inputs at around the 25,000 level.
| Development Stage | Monthly Calls | Average Input Tokens per Call | Average Output Tokens per Call | Total Monthly Input Tokens | Total Monthly Output Tokens |
|---|---|---|---|---|---|
| Prototyping & Initial Design | 150 | 3,000 | 1,500 | 450,000 | 225,000 |
| Core Logic Implementation | 450 | 25,000 | 2,000 | 11,250,000 | 900,000 |
| Refactoring & Code Review | 200 | 15,000 | 1,000 | 3,000,000 | 200,000 |
| Debugging & Error Tracing | 200 | 45,000 | 2,500 | 9,000,000 | 500,000 |
| Monthly Total | 1,000 | -- | -- | 23,700,000 | 1,825,000 |
When calculating API expenditures, you must factor in the difference in unit prices between input and output tokens. Because output tokens consume real-time computing resources, they are at least 3 to 5 times more expensive than input tokens. Total monthly expenditure is calculated using the following formula:
ext{Cost}*{ ext{monthly}} = sum*{m} left( rac{T_{ ext{input}, m}}{1,000,000} cdot P_{ ext{input}, m} + rac{T_{ ext{cache}, m}}{1,000,000} cdot P_{ ext{cache}, m} + rac{T_{ ext{output}, m}}{1,000,000} cdot P_{ ext{output}, m} ight)If you handle all tasks with a single model like Claude 3.5 Sonnet ($3.00 per 1M inputs, $15.00 per 1M outputs), based on monthly totals (approx. 23.7M inputs, approx. 1.825M outputs), it comes out to $98.48 (approx. 135,000 KRW). The problem arises when conversations grow long and context accumulates. If the input volume triples, monthly expenditures immediately exceed $250. Solo developers should set an upper AI spending limit at $100 to 50) and 80% ($80) alerts along with a 100% Hard Limit in the provider console to stay safe.
If you feed an entire project folder directly into a coding agent, auxiliary files like node_modules, dist/, and package-lock.json get included, causing 60% to 70% of your input tokens to vanish into thin air. It is better to use the CLI tool Repomix to selectively pack only the necessary business logic files.
json { "output": { "filePath": "repomix-output.xml", "style": "xml", "removeComments": true, "removeEmptyLines": true }, "ignore": { "useGitignore": true, "useDefaultPatterns": true, "customPatterns": [ "**/node_modules/<strong>", "</strong>/dist/<strong>", "</strong>/*.test.ts", "**/*.spec.ts", "<strong>/package-lock.json", "</strong>/yarn.lock", "**/*.public/<strong>", "</strong>/*.svg" ] } }
Apply prompt caching to system prompts and common library code. For the Anthropic Claude API, if you insert a cache_control breakpoint into a system prompt of at least 1,024 tokens, making a repeat request within 5 minutes of initial creation applies a 90% discounted cache read unit price of $0.30/1M.
json { "model": "claude-3-5-sonnet-20241022", "max_tokens": 2048, "system": [ { "type": "text", "text": "The user is a solo developer specializing in TypeScript and Next.js. Return only code tailored to precise interface specifications.", "cache_control": {"type": "ephemeral"} } ], "messages": [ { "role": "user", "content": "Common DB schema definition file contents..." } ] }
Methods for shortening prompt length can be summarized in three steps:
--compress option when running Repomix to strip out internal implementation details and extract only interfaces and function signatures.--remove-comments and --remove-empty-lines flags to chop away comments and whitespace, reducing token length by another 15% to 20%.There is no need to use expensive models for simple component placement or Tailwind CSS styling tasks. For frontend UI drafts, use Gemini 2.0 Flash ($0.10/1M input, 0.15/1M input, $0.60/1M output), which process long contexts affordably.
| Work Area | Detailed Task Content | Recommended Model | Reason for Selection |
|---|---|---|---|
| Frontend UI | Screenshot-based Tailwind CSS writing, HTML layout configuration | Gemini 2.0 Flash | Affordable multimodal processing and fast response speed |
| General Backend | RESTful API implementation, basic DB CRUD, input validation | DeepSeek V3 | Delivers solid performance at 1/10th the unit price of Claude |
| High-Difficulty Backend | Multi-transactions, concurrency control, complex refactoring, security audits | Claude 3.5 Sonnet | Pinpoint replacement when lower-tier models repeat errors |
Write general CRUD or one-off API logic using DeepSeek V3 ($0.25/1M input, $0.95/1M output), which has a unit price about one-tenth that of Claude 3.5 Sonnet. Switch to Claude 3.5 Sonnet only when errors cannot be resolved in tricky code, such as complex queries involving multiple tables or concurrency control transactions.
To minimize context waste between these two models, use the following connection workflow:
Public benchmarks measure performance via one-off queries, making them different from coding agent environments (like Aider). If a conversation exceeds 10 turns, the previous conversation history is transmitted every time, costing $1 to $2 just to fix a single typo.
If the agent falls into an autonomous correction loop of fixing code and re-executing after a test failure, $20 to $50 can vanish in minutes. To prevent this, three control criteria must be put in place:
| Verification Item | Detailed Review Content | Execution & Inspection Method |
|---|---|---|
| Prompt Isolation | Confirm ignored file settings and packing size | Measure packing token count in advance after applying .gitignore and repomix.config.json |
| Model Distribution | Assign models by task difficulty | Designate Gemini Flash for UI, DeepSeek V3 for CRUD, and Claude Sonnet for high-difficulty logic |
| Prompt Caching | Verify minimum token units and breakpoints | Confirm system prompts are 1,024+ tokens and insert cache_control |
| Spending Cap and Loop Limits | Configure safety circuit breakers | Limit agent autonomous repetitions to 5 or fewer, and set spending limits in the API provider console |
The key to using AI for coding is not blindly attaching the latest expensive models, but controlling your tools according to task characteristics. By estimating your budget with token formulas, trimming prompts with Repomix, and mixing models to suit the nature of the work, you can complete products without financial strain.