Anthropic bills a 1568px-wide image at a flat rate regardless of how much text is inside it. Dense content (code, JSON, tool output) packs ~3.1 chars per image-token vs ~1 char per text-token on real Claude Code traffic. pxpipe is a local proxy that exploits that gap: it rewrites the bulky middle of your conversation into compact PNGs before the request leaves your machine.

Running against real Claude Code sessions, the production log (13,709 requests) shows: a $100 total bill becomes ~$41 — full dollar math, input

cache writes at 1.25× + cache reads at 0.1× + output at 5×, including the ~6k small requests pxpipe correctly leaves untouched. On the large requests it actually compresses (7,756 of them), $100 of spend becomes ~$28. Quote the 59%; the 72% only applies to touched requests.

This is what the model sees instead of text:

~48k characters of system prompt + tool docs (this repo's own README, FINDINGS, and source), ≈25k tokens as text, ≈2.7k image tokens as this page. Produced by the real transformRequest pipeline: whitespace-minified, reflowed into full rows with ↵ marking original newlines, OCR instruction banner co-rendered on top. The model reads renders like this at 100/100 on a clean eval (see benchmarks).

Try it (30 seconds)

npx pxpipe-proxy                                  # proxy on 127.0.0.1:47821
ANTHROPIC_BASE_URL=http://localhost:47821 claude  # point Claude Code at it

Open http://127.0.0.1:47821/ for a live dashboard: tokens saved, per-session stats, every text→image conversion side by side, and a kill switch.

Nothing else changes. Responses stream normally; pxpipe only compresses the request (your context going up), never the model's output. Recent turns stay text; only older bulk history is imaged.

The honest part, read before relying on it

It is lossy. pxpipe is a gist tier, not a lossless store. In a needle-in-haystack eval, exact 12-char hex strings inside dense imaged content came back 0/15 on Opus and 3/4 on Fable 5, and the failure mode is silent confabulation: a plausible wrong value, not an error. Anything you need back byte-exact (IDs, hashes, secrets, exact numbers) must stay text. Recent turns do; a dedicated verbatim-risk guard is not built yet.

Does it break real work? Not in what we measured: a 10-instance SWE-bench Lite pilot (the easy subset) resolved 10/10 on both arms — pxpipe ON at $27 vs OFF at $54 token-equivalent, and a SWE-bench Pro single-pair smoke test (harder, long-horizon) resolved 1/1 both arms at -66% per-request. Small n, parity not superiority; details and caveats in the benchmarks below.

Savings are workload-dependent. It wins on token-dense content (~1 char/token: code, JSON, hashes) and loses money on sparse English prose (~3.5 chars/token). The built-in gate only images content where the math wins, calibrated against N=391 production rows.

Model scope: Fable 5 only (claude-fable-5), enforced in library and proxy. Opus 4.7/4.8 was the original scope but misread ~7% of renders (10200→9400), so it was disabled once Fable 5 hit 100/100 with identical image billing. Everything else passes through untouched.

Benchmarks (reproducible)

Measured with novel random-number problems the model cannot have memorized:

test	N	text	pxpipe (image)	tokens
novel arithmetic, `claude-fable-5`	100	100%	100%	−38%
novel arithmetic, `claude-opus-4-8`	100	100%	93%	−38%
gist recall A/B (decisions, values, paths, names, negations; with distractors; 15k-45k char sessions), Fable 5	98/arm	98/98	98/98	-
state tracking (value mutated 3x, final/first/count), Fable 5	18/arm	18/18	18/18	-
confabulation on never-stated facts (lower is better), Fable 5	16/arm	0/16	0/16	-
verbatim 12-char hex recall, dense render, Opus	15	15/15	0/15	-
verbatim 12-char hex recall, dense render, Fable 5	4	-	3/4	-

SWE-bench Lite pilot (end-to-end task quality)

10 SWE-bench Lite instances, Claude Code + Fable 5, paired runs through pxpipe ON vs OFF, graded with the official swebench Docker harness:

	pxpipe ON	OFF
resolved	10/10	10/10
request size vs own uncompressed body	−65%	±0

The −65% is per-request (count_tokens probe of each body before compression), so it has no turn-count confound. n=10/arm, Lite skews easy. Run totals, receipts, caveats: eval/swe-bench/.

A SWE-bench Pro pilot (harder, long-horizon tasks) is at n=1: both arms resolved 1/1 with −66% per-request, consistent with Lite and the production log. Receipts: eval/swe-bench-pro/.

_{We also ran GSM8K: 96% imaged. But GSM8K is in training data, so the model
recalls memorized answers through its own misreads, inflating the score, so we
lead with the clean novel-number eval instead. Reproduce:
eval/gsm8k/ · eval/needle-haystack/ ·
eval/gist-recall/ ·
full analysis in FINDINGS.md.}

How it works

tool_result string ──► wrap at 1568px-wide columns ──► pack ~5,000 chars/page ──► PNG[]

The proxy intercepts /v1/messages, rewrites eligible bulk history into image blocks, splices them back cache-friendly (static prefix preserved, so prompt caching keeps working), and forwards. Per-request events log to ~/.pxpipe/events.jsonl.

The economics: a 1568×1568 image costs ≈1,568 vision tokens and holds ≈5,000 readable chars (≈1,250 text tokens), so plain text is cheaper unless your text is token-dense. Claude Code transcripts are (observed 1.91 chars/token, N=391). The runtime estimator (estimateImageCount) plus a chars/token gate decides per-request; sparse prose is left as text.

Library use (no proxy)

import { renderTextToPngs, estimateImageCount } from "pxpipe";

const pngs = await renderTextToPngs(toolResultText);  // Buffer[], attach to the next user turn

renderTextToPngs(text: string, cols?: number, style?: RenderStyle): Promise<Buffer[]>
estimateImageCount(text: string, cols?: number): number   // gate yourself
wrapLines(text: string, cols: number, markerScale?: number): string[]

constant	value	meaning
`DENSE_CONTENT_CHARS_PER_IMAGE`	5 000	target chars per page
`READABLE_CHARS_PER_IMAGE`	50 000	hard ceiling per page
`DEFAULT_COLS`	313	column width
`MAX_HEIGHT_PX`	1 568	page height ceiling

Development

pnpm install && pnpm test     # 323 tests
pnpm run build                # regenerates dist/

Limitations

Lossy: see "the honest part" above. Verbatim recall from images is unreliable.
Render latency: encoding PNGs adds time to large requests before they leave (partly offset by the model ingesting fewer tokens). Responses stream normally.
ASCII/Latin-1 well tested; CJK works but conservatively.
node-canvas native dep on Node.
Fable 5 only.

License

MIT.

README.md Unescape Escape

pxpipe