A local proxy called pxpipe turns bulky context such as system prompts, tool docs, and older chat history into compact PNG images before they reach Fable 5. The model recovers the text through OCR while recent turns and precise values stay as plain text, producing reported bill reductions of roughly 60 percent on dense workloads by leaning on the cheaper vision-token pricing tier.
I noticed curious price scaling for this myself when working on visual memory last year. The cost to interpret images was not commensurate across methods and platforms.
~60% Fable cost cut by transparently turning the code into an image and having the model OCR it. WILD idea. also hilarious. https://github.com/teamchong/pxpipe
Dense text rendered as images consumes far fewer tokens than the equivalent string under Fable 5 rates, with GitHub examples showing 25 k text tokens shrinking to about 2.7 k image tokens and real session costs dropping from $42 to $6.
The approach is lossy by design, so hex strings, IDs, and other precise values can drop from full accuracy to 13 out of 15; the repo therefore routes those items away from the image path to avoid silent mistakes.
Some users praise the Fable AI hack of rendering code as images for OCR as brilliant and effective at cutting costs, while others dismiss token-based pricing as stupid or predict the approach will lead to higher image prices.
No Digg Deeper questions have been answered for this story yet.
@MiTypeScript Wow, that is very interesting. Does it actually work?
@EthosVentures https://www.seangoedecke.com/text-tokens-as-image-tokens more here, including a link to the white paper
@MiTypeScript Lol that makes no sense and just means images are mispriced and will soon increase unfortunately
~60% Fable cost cut by transparently turning the code into an image and having the model OCR it. WILD idea. also hilarious. https://github.com/teamchong/pxpipe
I don't understand this. If it is indeed that much cheaper and not just a mispricing, as the DS paper says it isn't, why aren't all the labs just doing this anyway in the backend to increase margins and cut prices?
~60% Fable cost cut by transparently turning the code into an image and having the model OCR it. WILD idea. also hilarious. https://github.com/teamchong/pxpipe
@dtdev_x maybe @sjgoedecke can tell us? https://www.seangoedecke.com/text-tokens-as-image-tokens was something that came up trying to understand why this works
@EthosVentures read the DeepSeek paper. it's more than a pricing hack, as bizarre as that may be
@phillpafford lol, for now. depends if it's cheaper because it's more efficient or cheaper because gaming the pricing model. if the latter.. probably a ToS endgame in sight.
@MiTypeScript @flyosity a picture is worth a thousand tokens
@MiTypeScript lmao someone plz benchmark this
@MiTypeScript This sounds like DeepSeek OCR
@MiTypeScript How does that even make sense
@MiTypeScript Also see https://blog.can.ac/2026/06/10/snapcompact/ by @_can1357. Works very well in practice!
@krishnanrohit because it changes the semantics from the pov of the model and there's a known eval phenomenon of "text prompts given as vlm images cause overall performance/reasoning degradation compared to naive text prompts"
@MiTypeScript another reason why i think charging for tokens is so stupid from a consumer perspective bc they really should be charging by actual energy consumption like Neuralwatt.
@MiTypeScript Which paper? Either the quality of the context from the ocr is lower, or the context is the same quality and therefore images are mispriced, no?
@MiTypeScript deepseek wrote about this months ago
@MiTypeScript CIE CHS and knowing that he is designed to cycle at night+ WBC dreaming , is catastrophic ! He's not solving medicine or making new genes Rn! Take out this crap from Claude! NCA (Neural Cellular Automata) Engine. Morphogenetic Self-Assembly + Liquid Nodes are enough!
@MiTypeScript i talk about this here
@MiTypeScript surely the model will struggle to attend to text in encoded image tokens compared to text directly? even if it is able to recall specific parts.. i would be surprised if you did not see a significant performance hit on how well it "perceives" this text
