LLM token-waste audit corpus and public audit statistics, 2026
Precision and recall of the static token-waste auditor on the golden corpus, plus aggregate statistics from public repository audits.
The golden audit corpus has 13 cases; the static auditor scores precision 100%, recall 100% and a 0% false-positive rate on clean files as of 2026-10-26. Across 75 public repository audits the median estimated waste is 0% and 45.33% carry at least one finding.
· Source: TokenZero pricing table, benchmark result files and public audit ledger; CC BY 4.0 with attribution
At a glance
- Corpus cases
- 13
- Precision
- 100% (computed)
- Recall
- 100% (computed)
- F1
- 1 (computed)
- Clean-file false positives
- 0% (computed)
- Public audits
- 75
- Median estimated waste
- 0% (computed)
- Audits with findings
- 45.33% (computed)
LLM token-waste audit corpus and public audit statistics, 2026 (first 13 of 13 rows; full dataset in the downloads)
| case | path | clean_file | expected_findings | found_findings | waste_pct | pass |
|---|---|---|---|---|---|---|
| py-openai-classifier-uncached | app/triage.py | 0 | static-prompt|cascade|repeat-cache | cascade|static-prompt|repeat-cache | 70 | 1 |
| ts-anthropic-system-no-breakpoint | src/assistant.ts | 0 | static-prompt | static-prompt | 70 | 1 |
| py-loop-per-item | jobs/enrich.py | 0 | loop | loop | 15 | 1 |
| ts-vercel-ai-json-extraction-frontier | lib/extract.ts | 0 | cascade | cascade | 70 | 1 |
| py-gemini-static-prefix-sentiment | svc/sentiment.py | 0 | static-prompt|repeat-cache | static-prompt|repeat-cache | 12 | 1 |
| py-langchain-translate-frontier | pipeline/translate.py | 0 | cascade | cascade | 70 | 1 |
| py-retiring-model | svc/digest.py | 0 | lifecycle | lifecycle | 0 | 1 |
| lib-sdk-wrapper | src/agents/models/openai_chat.py | 1 | 0 | 1 | ||
| ts-anthropic-already-cached | src/cached.ts | 1 | 0 | 1 | ||
| py-already-cheap-model | app/tag.py | 1 | 0 | 1 | ||
| docs-model-names-in-prose | docs/models.md | 1 | 0 | 1 | ||
| py-batched-loop | jobs/batch_enrich.py | 1 | 0 | 1 | ||
| ts-openai-cache-options-set | src/summarize.ts | 1 | 0 | 1 |
Method and licence
Computed at request time from the same data the gateway runs on; the JSON summary carries every number on this page. Reuse under CC BY 4.0 with a link to https://tokenzero.sh/reports/llm-token-waste-audit-corpus-2026.
Questions people ask
What does the auditor look for?
Uncached static system prompts, hard-coded expensive or retired model ids, padded tool schemas, reasoning-heavy models on simple tasks and missing cache keys; findings carry file and line. It runs locally in the CLI, the MCP server and the GitHub Action; nothing is uploaded.
How accurate is it?
Precision 100% and recall 100% on the 13-case golden corpus, with 0% false positives on clean files; the corpus and scorer ship in the public repository.
Are the public audit statistics real repositories?
Yes: every row behind the summary is a hosted audit of a public GitHub repository that someone ran, listed on /audits with its receipt page.