compress-on-input Performance Report

Real-world compression statistics from 11 days of daily use with Claude Code

March 6–16, 2026 3,257 tool calls 0 errors

93.4%

Overall reduction

9.9M

Tokens saved

$148

Saved (Opus pricing)

15ms

Median latency

Compression by Content Type

Screenshots → OCR

99.8%

181 screenshots processed via Apple Vision OCR. Base64 image data replaced with extracted text. Avg OCR time: 735ms.

~210k chars → ~425 chars

JSON / DB Results

98.6%

60 large JSON responses collapsed to schema + 3 sample rows. Avg time: 1ms.

~9.6k chars → ~137 chars

Large Text

50.0%

1 large text document truncated via BM25-ranked chunking. Head + relevant middle + tail preserved.

533k chars → 266k chars

DOM Snapshots

24%

1,761 accessibility tree snapshots cleaned up. Most frequent type. Ref mapping preserved for clickability.

~1.8k chars → ~1.4k chars

Token Savings Breakdown

Where the 9.9 million saved tokens came from:

Screenshots → OCR

9.5M tokens

95.9%

DOM snapshots

190k tokens

1.9%

JSON / DB

142k tokens

1.4%

Large text

67k tokens

0.7%

Per-Tool Breakdown

Tool	Compressed	Passed Through	Total Before	Total After	Reduction	Avg Latency
browser_take_screenshot	181	—	38.0M	77k	99.8%	735ms
browser_click	604	39	701k	557k	20.6%	19ms
browser_navigate	458	28	560k	422k	24.6%	28ms
browser_snapshot	383	36	1.2M	863k	27.1%	20ms
browser_wait_for	205	40	638k	495k	22.3%	21ms
browser_evaluate	3	297	2.6k	2.6k	0.3%	16ms
browser_run_code	39	272	27k	27k	0.5%	26ms
db_query	36	—	534k	4.9k	99.1%	1ms
notion-search	4	—	19.5k	1.5k	92.3%	22ms
query (D1)	12	—	18.6k	1.1k	94.0%	10ms

DOM Snapshot Compression Distribution

1,761 DOM snapshots — how much each one was reduced:

<5% reduction

230

13.1%

5–15% reduction

426

24.2%

15–30% reduction

794

45.1%

30–50% reduction

301

17.1%

>50% reduction

0.6%

Screenshot OCR Quality

181 screenshots processed — quality of extracted text:

Rich (>500 chars)

34.8%

Medium (200–500)

31.5%

Short (<200 chars)

33.7%

Empty (<50 chars)

0.0%

Zero empty OCR results — every screenshot produced usable text. "Short" results are typically screenshots with minimal text content (dialogs, confirmations, loading states).

Daily Activity

Compressed events per day and chars saved:

Mar 678910 111213141516

Latency Profile

By operation type

DOM cleanup

20ms

JSON collapse

1ms

Text truncate

58ms

OCR (cached)

735ms

OCR (first call)

~3s

Overall percentiles

Median (P50)

15ms

P95

559ms

Max

6.6s

Mean

86ms

28 operations exceeded 1 second — all screenshot OCR. The 6.6s max was the first-ever OCR call (Swift binary compilation).

Beyond Size Reduction — Quality of Life

Compression isn't just about tokens. It changes how Claude interacts with data.

Screenshots become searchable text

Instead of a 210k-char base64 blob that Claude processes as an image, it receives semantic text. Claude can search, compare, quote, and reason about page content across multiple screenshots.

Before: data:image/png;base64,iVBORw0KGgo... (210k chars)

After: Dashboard — Revenue: $45,230 (+12.3%)
Active Users: 1,847 | Conversion: 3.2%
Top Campaign: "Spring Sale" — 234 signups

DB queries return schema, not noise

500-row query results become schema + 3 examples. For understanding structure, writing follow-up queries, or debugging — the schema is more useful than scrolling through raw data.

Before: [{"id":1,"name":"Alice","email":"a@..."}, {"id":2,...}, ... × 500 rows]

After: Array[500] of {id, name, email, created_at, role}
Sample: {id: 1, name: "Alice", email: "a@b.com", role: "admin"}
Sample: {id: 2, name: "Bob", ...}

DOM stays clickable, less noisy

Inline [ref=eNN] markers are extracted into a compact mapping table. Claude still clicks elements by ref, but the text reads more naturally.

Before: - button [ref=e45] "Submit" [role="button"]
- div [ref=e46] [role="generic"]
- div [ref=e47] [role="none"]

After: - button "Submit"
Refs: Submit→e45

Smart passthrough prevents data loss

1,254 small results passed through untouched. browser_evaluate, browser_fill_form, browser_type — outputs that are already compact and shouldn't be touched.

Top passthrough tools:
browser_evaluate: 297 ✓
browser_run_code: 272 ✓
browser_press_key: 72 ✓
browser_fill_form: 61 ✓
Small output → no compression needed

Reliability

Errors in 3,257 events

Negative compressions (max -0.2%)

Slow ops (>1s) — all OCR

100%

Uptime (fail-safe design)

The fail-safe design ensures the hook never makes things worse: if compression fails or increases size, the original content is returned unchanged. In 11 days of continuous use, there were zero instances of data loss or corruption.

Top 10 Biggest Single Saves

#	Tool	Before	After	Saved	Reduction
1	browser_take_screenshot	471,037	1,405	469,632	99.7%
2	browser_take_screenshot	430,814	1,269	429,545	99.7%
3	browser_take_screenshot	430,810	1,265	429,545	99.7%
4	browser_take_screenshot	375,522	960	374,562	99.7%
5	browser_take_screenshot	375,517	955	374,562	99.7%
6	browser_take_screenshot	352,010	851	351,159	99.8%
7	browser_take_screenshot	349,034	791	348,243	99.8%
8	browser_take_screenshot	337,322	1,014	336,308	99.7%
9	browser_take_screenshot	335,167	712	334,455	99.8%
10	browser_take_screenshot	303,117	760	302,357	99.7%

All top saves are screenshots — a single full-page screenshot can consume ~117k tokens. OCR replaces that with ~300 tokens of text, saving the equivalent of a medium-sized conversation.