Your personal operating system for career transition.

Implement Token Counting and Cost Tracking

Token usage is the billing unit for every major LLM API. Without explicit tracking, AI features run with no cost visibility until the monthly invoice arrives. Engineers who track token usage per request can answer "which feature costs the most?", "is this prompt getting more expensive over time?", and "what is the P95 cost per user session?"

What to build

Implement a CostTracker class that records token usage across multiple LLM calls and computes cost summaries:

1. record(model: str, input_tokens: int, output_tokens: int, feature: str) method — records one LLM call with its token usage and the product feature that made the call.
2. total_cost_usd() method — returns the total USD cost across all recorded calls.
3. cost_by_feature() method — returns a dict mapping feature name to total USD cost.
4. cost_by_model() method — returns a dict mapping model name to total USD cost.
5. summary() method — returns a dict with total_calls, total_input_tokens, total_output_tokens, total_cost_usd, cost_by_feature, cost_by_model.

Also implement MODEL_COSTS: dict[str, dict] — a lookup table mapping model name to {"input_per_1k": float, "output_per_1k": float}. Include at minimum:

• "gpt-4o": input $0.0025/1k, output $0.01/1k
• "gpt-4o-mini": input $0.00015/1k, output $0.0006/1k
• "claude-3-5-sonnet": input $0.003/1k, output $0.015/1k

Why this matters

Token tracking is required for cost attribution in multi-feature products, budget alerts before hitting billing limits, and deciding which model tier to use based on actual usage patterns. A tracker that is easy to integrate (one .record() call per LLM response) gets adopted; a complex one gets bypassed.

Constraints

• Raise ValueError for unknown model names.
• All costs should be computed from the MODEL_COSTS table, not hardcoded.
• The summary() method should return costs rounded to 6 decimal places.