Pixels for Programs? A Cross-Provider Case Study of Input-Token Accounting for Source Code as Text and Images

2026-07-27 · Source: cs.SE updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Data Science & Analytics · Depth: Expert, extended

Summary

The article presents a case study on input-token accounting for source code sent to commercial vision-language model APIs as text versus compact images. The study benchmarked requests across five programming languages, nine source lengths from 20 to 2,000 lines, and 15 model aliases exposed by Anthropic, OpenAI, and Google Vertex AI (Gemini). Across 675 complete text/image pairs, aggregate image-to-text ratios were 0.135 for Anthropic, 0.194 for OpenAI, and 0.242 for Gemini, corresponding to reported input-token reductions of 86.5%, 80.6%, and 75.8% respectively. However, these aggregates conceal varying break-even behaviors: Anthropic and OpenAI images consistently received lower counts, while Gemini images required 6.95 times more tokens at 20 lines and only crossed below text at 200 lines. A targeted audit also revealed non-monotonic Gemini image accounting across page boundaries. The study provides a reproducible protocol and artifact but does not measure semantic fidelity, task accuracy, latency, monetary cost, or coding-agent efficiency.

Key takeaway

For Machine Learning Engineers optimizing API costs for code processing, you should implement conditional routing for source code inputs. For short contexts, raw text is often more cost-effective, especially with Gemini. For longer, read-mostly contexts, visual representation can reduce reported tokens, but you must validate fidelity and monitor provider-specific break-even points and page-sensitive accounting. Always compare against textual compression and retrieval methods.

Key insights

Sending source code as compact images to LLM APIs can significantly reduce reported input tokens, but with provider-specific break-even points.

Principles

• Image token accounting varies significantly by provider.
• Aggregate ratios can hide fixed-cost behavior.
• Token reduction does not imply cost or quality.

Method

The study used a two-stage transform: replacing indentation with markers, then rendering to PNG pages. It collected provider-reported input tokens for raw text and compact images across varied languages and lengths.

In practice

• Measure break-even points for each provider.
• Monitor image dimensions and page counts.
• Compare visual conversion with textual compression.

Topics

• Vision-Language Models
• Token Accounting
• Source Code Compression
• Multimodal APIs
• Empirical Software Engineering
• API Cost Optimization

Code references

• ron-42/code-image-token-accounting

Best for: AI Architect, AI Engineer, CTO, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• A Picture Costs a Thousand Words — Literally — Vision Transformers' quadratic attention cost for image tokens drives high expense, despite most visual tokens being redundant.
• The HackerNoon Newsletter: Why Cost Per Token Is the Wrong AI Metric (7/10/2026) — Total cost per successful task, not cost per token, determines AI model value.
• Only about 1% of the tokens Thariq generates go into production code — AI-driven content generation reduces documentation costs, enabling focus on idea quality and flexible, just-in-time practices.
• Token Reduction Is Not Cost Reduction — Token reduction does not equate to cost reduction for API-based coding agents due to cache pricing and trajectory effects.
• Are All Tokens Necessary for Visual Place Recognition? An Empirical Study of Token Reduction for Efficient Inference — Token reduction significantly boosts VPR inference efficiency with minimal accuracy loss, enabling real-time deployment for vision transformers.
• Token Economics: Why LLM Output Tokens Cost More Than Input Tokens — LLM output tokens cost more due to sequential generation and increasing memory bandwidth demands.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.SE updates on arXiv.org.