Stop Paying Full Price: How This Genius Turned Claude Code Prompts Into Images to Cut Costs by 60%
Forget LLM token inflation. An enthusiast just hacked the Anthropic pricing model by forcing Claude Code to 'read' pictures instead of text. It is absurd, slightly broken, and absolutely brilliant engineering.
The tool, dubbed pxpipe, works as a local proxy that intercepts API calls to Anthropic. Since the pricing for images is fixed based on dimensions, stuffing dense code, JSON, or documentation into a 1568-pixel wide PNG is significantly cheaper than paying for individual text tokens. The proxy selectively renders heavy context into images while keeping fresh conversational input as standard text.
This method achieves massive savings because high-density content, like source code or logs, takes up far fewer tokens in image form than in plain text. Testing reveals that while the model handles logic perfectly, it struggles with precision, often hallucinating when asked to extract exact hex strings or identifiers from the 'image' context. Claude 3.5 Sonnet is the current sweet spot for this setup, as older Claude Opus variants struggle with OCR-style errors, occasionally mangling numbers.
While pxpipe delivers on efficiency, it isn't a silver bullet for perfectionists. It maintains high performance on benchmarks like SWE-bench, but the risk of subtle data corruption remains high for tasks requiring byte-for-byte accuracy. It is a classic case of trading absolute reliability for a lighter wallet, proving once again that in the current AI gold rush, the smartest move is often to outsmart the billing algorithm itself.
Source: pxpipe
Comments
This is where the magic happens: AI reads your discussion and rewrites the article based on the most interesting comments. Each strong comment adds points to the meter below. Once the meter is full, the article updates live — no page reload needed.