Written with AI assistance from the source video. The video is the authority.
Log in to leave a comment
No posts yet
For freelancers or small teams working with a monthly development budget of under 500,000 KRW, Anthropic's recent policy changes are quite a headache. Using third-party tools has become restricted, and any remaining monthly credits do not roll over to the next month and simply disappear. However, if you rely blindly on official clients, just a few coding sessions can easily push your API bill past hundreds of thousands of krw.
The problem won't be solved just by changing tools. You need to overhaul the prompt-exchange structure itself to protect your bank balance.
Agent-based coding tools send tens of thousands of tokens of context all at once with a single request. According to Anthropic's official pricing table, the base input cost for Claude 3.5 Sonnet is $3.00 per million tokens, and output is $15.00. On the other hand, applying prompt caching drops the input unit price to $0.30. That's a 90% difference. The moment a third-party extension injects a timestamp or session ID at the very top of the system prompt, all of this cache is broken, and you end up paying $3.00 every time.
First, open your logs from the past three months and calculate the ratio of cache_read_input_tokens to input tokens.
git diff as context.cache_control: {"type": "ephemeral"}.Simply moving dynamic variables to the bottom of the prompt and caching static text can eliminate nearly half of your background token waste.
There is no need to write every line of code with Sonnet or Opus. Calling cloud frontier models for tab autocomplete is a clear waste.
| Task Classification | Called Model | Execution Location | Cost Level |
|---|---|---|---|
| Architecture design and large-scale refactoring | Claude 3.5 Sonnet | Cloud API | Standard rate (100%) |
| Simple utility functions and unit tests | Claude 3.5 Haiku, DeepSeek-V3 | Cloud API | Around 20% to 30% |
| Real-time tab autocomplete and boilerplate | Ollama (Qwen2.5-Coder 14B) | Local On-Device | $0.00 |
Spin up LiteLLM on your local machine to act as a traffic gateway. Write model routing rules in config.yaml.
`yaml
model_list:
`
Specify export ANTHROPIC_BASE_URL="http://localhost:4000" in your terminal and connect Continue.dev or Claude Code. Real-time autocompletion is handled by the local GPU's Qwen model, while complex design questions are forwarded to Sonnet.
Simultaneously work on reducing output tokens. If you write system prompts to omit all unnecessary introductions, greetings, and code explanations and return only code blocks, high-unit-price output token spending will decrease significantly.
Anthropic's monthly credits simply evaporate once the expiration date passes. To avoid throwing away the remaining balance, you should run the Anthropic Batch API 3 days before expiration.
Batch API cuts all token prices by 50% under the condition of asynchronous return within 24 hours. Combine this with prompt caching, and you can process tasks at about 5% of the original price. Gather tasks that don't need immediate real-time answers and submit them all at once.
By bundling these tasks into a JSONL file and pushing them to the batch endpoint, you can turn credits that were about to be wasted into test code and documentation assets.
To prevent excessive charges, turn off Auto-Recharge in the Anthropic console's billing settings and manually pay around $150 at the beginning of the month.
After binding the direct call limit to $150 in the console's Spend Limits menu, add fallback settings to the LiteLLM proxy.
`yaml
router_settings:
fallbacks:
- claude-3-5-sonnet-20241022: ["ollama/qwen2.5-coder:14b"]
`
Even if credits run out and a 429 error occurs, the local Ollama immediately takes over the calls. If working at a team level, share .continue/config.json at the root of the Git repository and set the monthly limit for each team member's virtual key to $50 on the LiteLLM admin page. Since only the traffic of developers who exceed the limit is switched to the local model, the team's overall spending will never exceed the budget.