Token Tracker Methodology
Every waste signal is deterministic — no AI inference, no black-box scoring. This page defines exactly when each signal fires, how the waste estimate is computed, and what to do about it.
How costs are recomputed
Your CSV export contains token counts and sometimes provider-reported costs. The tracker ignores the provider's reported cost and recomputes spend using the Cromus verified model-price registry (the same registry used by the Waste Check tool). This ensures consistent, comparable numbers regardless of which provider exported the data.
Cost formula: (input_tokens / 1,000,000) × input_price + (output_tokens / 1,000,000) × output_price
| Provider | Token columns used | Date column | Notes |
|---|---|---|---|
| OpenAI | n_context_tokens_total, n_generated_tokens_total | timestamp | Cost recomputed from registry |
| Anthropic | input_tokens, output_tokens, cache_read_input_tokens, cache_creation_input_tokens | date | Cache tokens summed for cache rate signal |
| Google AI Studio | input_token_count, output_token_count | date | model_id or model used for registry lookup |
| OpenRouter | prompt_tokens, completion_tokens | created_at | total_cost column used as detection hint |
Waste signal definitions
Low Cache Utilization
The tracker measures cache hit rate as cached_tokens / (input_tokens + cached_tokens). When caching is clearly in use (≥20% of rows report cached tokens) but the hit rate is below 10%, repeated prompts and system prompts are paying full input-token prices for content that could be served from the cache at 10–25% of the original cost.
Enable prompt caching (Anthropic) or prefix caching (OpenAI) for system prompts and shared context. Target a cache hit rate above 25%.
High Output-to-Input Ratio
Output tokens are priced at 3–8× the rate of input tokens for most frontier models. This signal computes the per-row (day + model) ratio of output to input tokens and averages across all rows. A mean ratio above 3.0 indicates the model is generating far more tokens than it receives — common causes include verbose chain-of-thought, missing max_tokens limits, and large structured payloads that could be templated.
Add explicit max_tokens limits, use JSON mode to reduce prose overhead, and instruct the model to be concise in production prompts. Separate chain-of-thought reasoning from production output.
Premium Model Concentration
Quality-tier models (GPT-5.6, Claude Opus 5, Claude Fable 5, Gemini 3.5 Flash, and equivalents — excluding mini/haiku/flash-lite/nano variants) excel at complex reasoning and generation but are dramatically over-specified for classification, extraction, summarization, or retrieval tasks. This signal fires when a single quality-tier model dominates over 70% of total spend, indicating most requests are routed to the premium tier by default rather than by task complexity.
Audit your task mix by workflow. Route classification, summarization, formatting, and extraction calls to GPT-5.4 nano, Claude Haiku 5, or Gemini 3.5 Flash-Lite. Reserve quality-tier models for generation, multi-step reasoning, and customer-facing outputs.
7-Day Spend Spike
The tracker compares the most recent 7 calendar days (ending on the latest date in your dataset) against the preceding 7 days. A 40% or greater jump in spend most commonly signals unbounded retry loops, a newly deployed workflow without token budgets, or viral traffic hitting an unguarded workflow. ISO week grouping is intentionally avoided to preserve sensitivity at month boundaries.
Set per-workflow token budgets, add circuit breakers on retry logic, and configure spend alerts at 80% of your expected weekly baseline. Review new workflow deployments that coincide with the spike date.
High Cost on Short Interactions
Quality-tier rows where input_tokens < 500 are identified as short-context interactions — typically classification, labeling, scoring, boolean extraction, or terse summaries. If those rows collectively account for more than 15% of total spend, cheaper models (GPT-5.4 nano, Claude Haiku 5, Gemini 3.5 Flash-Lite) would handle them with comparable accuracy at 5–15× lower cost per call.
Profile your most common prompt patterns. If output is consistently short, route those calls to a lightweight model. Use quality-tier models only when the task genuinely requires advanced reasoning, long-form generation, or nuanced judgment.
Privacy & data handling
- • The Token Tracker requires sign-in. Uploaded usage is stored in the signed-in user's account; anonymous Token Tracker uploads are not available.
- • Authenticated uploads are stored per-account and are only visible to the authenticated user.
- • Duplicate rows (same user + date + model + provider) are skipped on re-upload — data is never double-counted.
- • Token counts and computed costs are stored; raw CSV rows are discarded after processing.
- • No row data is used for training, benchmarking, or any purpose other than displaying your tracker.