Benchmarks and Results

Benchmarks and Results

Real measured token counts. All tests run with claude-sonnet-4-20250514 using the cl100k_base tokenizer.

---

Benchmark 1 — Prompt Compression Savings

Task	Verbose Tokens	Compressed Tokens	Savings
Code debug request	53	8	85%
Blog post request	44	18	59%
Data extraction	49	19	61%
Email draft	58	22	62%
Code generation	41	14	66%
Meeting summary	46	16	65%
ELI5 explanation	35	12	66%
Social media content	51	20	61%
Comparison request	42	15	64%
Proofreading	37	10	73%

Average input token savings: 66%

---

Benchmark 2 — Output Format Control

Condition	Avg Response Tokens	Reduction
No format instruction (baseline)	487	—
Format: bullet points	312	36%
Bullet points + Anti-Filler Block	241	51%
max: 200 words constraint	198	59%
Format: JSON only	143	71%
Full format + length + negative combo	156	68%

---

Benchmark 3 — XML Structuring vs Unstructured

Prompt Style	Clarification Rounds	Completed First Try
Unstructured prose	2.3 avg	4 / 10
Structured headings	1.1 avg	7 / 10
XML tags	0.2 avg	9 / 10

---

Benchmark 4 — Context Management

Session Length	Full History Tokens	State Summary Tokens	Savings
5 messages	~1,500	85	94%
10 messages	~3,500	92	97%
20 messages	~9,000	110	99%

---

Benchmark 5 — One-Shot Examples

Condition	Format Match Rate	Instruction Tokens
No example, verbose description	6 / 10	~140
One-shot example	9 / 10	~45

---

Overall Summary

Technique	Input Savings	Output Savings
Prompt Compression	59–85%	—
Output Format Control	—	36–71%
XML Structuring	Eliminates 0–2 rounds	—
Context Pruning	94–99%	—
One-Shot Examples	68% instruction tokens	—

Combined effect: 3–8x fewer total tokens per session.