выпускать: 2026/11/21 17:15 читать: 0
Оригинальный автор:AICodeKing
Первоисточник:https://www.youtube.com/embed/aFmu4rLgH8U
Join this channel to get access to perks: https://www.youtube.com/channel/@aicodeking/join In this video, I'll take a look at pxpipe, an open-source local proxy that cuts coding agent input tokens by rendering bulky text context into dense 5x8 pixel page images. I'll explain why image token math beats raw text on dense code and JSON, what the proxy compresses versus leaves alone, and the headline end-to-end bill savings, session demo, and SWE-bench results. Next, I'll break down the part that matters most—which models can actually read 5x8 pixel text cleanly like Gemini 4 and Gemini Flash versus gist readers like Opus 5.5 or models you should not image at all—plus silent confabulations and the text factsheet fix. Finally, I'll show how to run npx pxpipe-proxy, put Claude Code, the Antigravity CLI, OpenCode, Codex, and omp behind pxpipe warp, use the dashboard, offline export, and pxpipe stats, and decide whether it's worth using. -- Key Takeaways: 🚀 pxpipe is an MIT-licensed local proxy on GitHub with around 7,600 stars that cuts coding agent input tokens by turning bulky context text into dense 5x8 pixel PNG pages. 📊 Because image cost is set by pixel size and one image token holds about 3.1 characters of dense code or JSON versus about 1 character per text token on Claude Code traffic, 48,000 characters of system prompt and tool docs drop from roughly 25,000 text tokens to about 2,700 image tokens. ⚡ It only compresses heavy parts of the request—the system prompt and tool docs, tool results above roughly 6,000 characters, and older chat history—while leaving your messages, recent turns, sparse prose, and streamed text responses untouched. 💸 On the original Claude Fable 5 from June, the project measured a 59% to 70% lower end-to-end bill including output tokens and cache reads, and its side-by-side Claude Code demo dropped from about $42 at 96% full context to about $6 at 73,000 out of 1 million tokens. ✅ On SWE-bench Lite both arms solved 10 out of 10 tasks with 65% less sent per request, and on SWE-bench Pro pxpipe got 14 out of 19 versus 15 out of 19 for plain text at 60% less, while every request logs true savings against the free count_tokens endpoint. 🎯 On the 15-hex-ID reading test, Gemini 4 Argon gets all 15, Gemini 3.6 to 3.8 Flash get 14, and original Fable 5 gets 13, whereas Opus 5.5 drops to 3 and Sonnet 5.5 gets 2, though 2 pixels of extra row spacing recover Fable 5.1 and 5.2 to 14 and 13. ⚠️ Models are sorted into full readers (Gemini 4, Gemini Flash, and original Fable 5), gist readers (Opus 5.5 and Fable 5.1 and 5.2), and models you should not image at all, including Haiku 5.5, Sonnet 5.5, GPT 6 Astra, Kimi K3, and DeepSeek V4.1 Flash. 🔍 Because model vision is not OCR, unclear pixels trigger silent confabulations rather than errors, so pxpipe attaches a small text factsheet of exact IDs and numbers next to each image to prevent glyph mix-ups such as 0 and 8. 💻 Start the proxy and dashboard on port 47821 with npx pxpipe-proxy, wrap Claude Code, the Antigravity CLI, OpenCode, Codex, or omp with pxpipe warp without changing base URLs, let subagents on unlisted models like Sonnet 4.6 stay plain text, or export pages offline and run pxpipe stats while keeping Anthropic's prompt cache boundary intact. 👍 Verdict: really worth trying on Gemini 4, Gemini Flash, or the original Fable 5—and useful on Opus 5.5 for normal tool-backed coding—just expect minor PNG encoding latency, skip it on Sonnet, Haiku, and open models, and watch the tier list across releases. -- Timestamps: 0:00 - Turning bulky context into images with pxpipe 0:24 - Why image tokens cost less than dense text 1:09 - How the local proxy works 1:27 - What gets imaged and the profitability gate 3:33 - Headline numbers and the $42 vs $6 session demo 4:07 - SWE-bench Lite and SWE-bench Pro 4:33 - Measuring real savings with count_tokens 4:45 - Which models can read 5x8 pixel text 5:26 - Model tiers full readers, gist readers and do-not-image 5:53 - Silent confabulations and the 0-vs-8 factsheet 6:41 - Running npx pxpipe-proxy on port 47821 6:53 - Putting agents behind pxpipe warp 7:30 - Dashboard, model chips and the subagent escape hatch 8:09 - Offline export for folders, stdin and git diffs 8:35 - Offline stats report with pxpipe stats 8:47 - API formats, prompt caching and the AI-written README 9:21 - Should you use it?
CRYPTO_BEN
2026-11-23 01:31
子敬.轮回珠宝
2026-11-23 01:31
kumis patiler
2026-11-23 01:31
Ha Ms
2026-11-23 01:31
CryptoMetric
2026-11-23 01:31
劇翻天 Drama Go
2026-11-22 16:36
CryptoMason
2026-11-22 16:22
Lý Tử Ly
2026-11-22 14:42
Pidoku
2026-11-22 14:15
Выберите валюту
US Dollar
USD
Chinese Yuan
CNY
Japanese Yen
JPY
South Korean Won
KRW
New Taiwan Dollar
TWD
Canadian Dollar
CAD
Euro
EUR
Pound Sterling
GBP
Danish Krone
DKK
Hong Kong Dollar
HKD
Australian Dollar
AUD
Brazilian Real
BRL
Swiss Franc
CHF
Chilean Peso
CLP
Czech Koruna KČ
CZK
Singapore Dollar
SGD
Indian Rupee
INR
Saudi Riyal
SAR
Vietnamese Dong
VND
Thai Baht
THB
Выберите валюту
US Dollar
USD-$
Chinese Yuan
CNY-¥
Japanese Yen
JPY-¥
South Korean Won
KRW -₩
New Taiwan Dollar
TWD-NT$
Canadian Dollar
CAD-$
Euro
EUR - €
Pound Sterling
GBP-£
Danish Krone
DKK-KR
Hong Kong Dollar
HKD- $
Australian Dollar
AUD-$
Brazilian Real
BRL -R$
Swiss Franc
CHF -FR
Chilean Peso
CLP-$
Czech Koruna KČ
CZK -KČ
Singapore Dollar
SGD-S$
Indian Rupee
INR -₹
Saudi Riyal
SAR -SAR
Vietnamese Dong
VND-₫
Thai Baht
THB -฿