Snapcompact renders context as pixel-font bitmaps, cutting LLM input costs ~2x
brandon_galang · x · 2026-09-15
Can Bölük published Snapcompact: when the context window fills up, render the whole context into dense pixel-font bitmaps and pass them back as images. A 1568×1568 PNG holds 40k chars (10k text tokens) billed as only 3,279 image tokens — roughly 1/3 the input price — and models read it back near-verbatim.
Benchmarks:
- Context carried: 170K tokens → 46K (÷3.7)
- Claude Fable 5 and Claude Opus 4.8: ÷2.0 total cost at F1 parity; GPT-5.5 ÷1.8 (−12pt F1); Gemini 3.5 Flash ÷2.9 (−6pt F1)
- Beats handoff and traditional summary compaction, which abstained heavily in tests
Free, local, instant — a practical cost optimization for long agent runs.
More from coding & agent
- How Much Code Is Throwaway Now? Engineers Rethink Project Lifespans in the AI Era — rakyll · 2026-09-15
- Translation MCP team: the hard part wasn't the server, it was getting the model to call it — Pure-Ad7786 · 2026-09-15
- Developer runs Claude inside Siri on macOS 27 via Apple's hidden model-provider support, open-sourced — Teknium · 2026-09-15
- OpenResearch tops GitHub trending, turns Claude Code and Codex into research agents — TheMoonMidas · 2026-09-15
- CoreSpeed Launches: One MCP Endpoint for Tools, Memory and Agent Budgets — Kind-Atmosphere9655 · 2026-09-15
- GitHub issue calls out Copilot CLI subagent latency: minutes per review round trip — tjgreen42 · 2026-09-15