Skip to content

Claims Library Entry

Token maxing is what you do when you can't budget

The HTML vs Markdown debate for AI-generated artifacts is really a symptom of a deeper problem: teams defaulting to maximum token usage without naming their constraints. The article argues for token budgeting—naming a primary constraint and letting format and model choices follow—backed by evidence from ETH Zurich, Anthropic, and Stanford/MIT studies.

Published May 12, 2026 by Kamil Banc

AI StrategyROI & MeasurementImplementation

Lead claim

Token maxing is what you do when you can't budget—name your constraint, or every AI choice is just a vibe.

Atomic Claims

What this article supports

Claim 1 · Source summary

Anthropic Engineer Backs HTML

An Anthropic engineer argued HTML should replace Markdown as the default format for AI-generated artifacts.

Claim 2 · Source summary

Mintlify Chooses Markdown

Mintlify reported a thirty times reduction in token usage when serving documentation as clean Markdown.

Claim 3 · Source summary

Context Files Backfire

An ETH Zurich study found LLM-generated context files cut task success rates by three percent.

Claim 4 · Source summary

Context Rot Is Universal

Chroma tested eighteen frontier models and found performance degrades at every context length increment.

Claim 5 · Kamil's interpretation

Maxing Without Constraint Is Defaulting

Token maxing without naming a primary constraint is defaulting dressed up as strategy, not deliberate choice.

Evidence

Context behind the claims

Quote

"If you can't name the trade in one sentence, you're not maxing. You're defaulting and dressing it up."

Key statistics

Up to 30x variance in token spend

Bai et al. at Stanford and MIT found running the same agent on the same task produces a thirty-fold range in cost, with models predicting their own usage at correlations no higher than 0.39.

3% lower success, 20%+ higher cost

ETH Zurich's SRI Lab study of 438 real-world coding tasks found LLM-generated AGENTS.md files cut task success rates by three percent while raising costs by over twenty percent.

80% fewer tokens served as Markdown

Cloudflare measured an eighty percent token reduction when serving the same blog page as Markdown instead of HTML, three weeks after Mintlify reported a thirty times reduction.

15x token multiplier for multi-agent research

Anthropic's multi-agent research system burns fifteen times more tokens than chat, a cost the team accepted deliberately after naming breadth-first parallelism as its primary constraint.

Supporting context

The article grounds its argument in peer-reviewed and industry research rather than ideology, citing controlled studies from ETH Zurich, Stanford and MIT, and Chroma's testing of eighteen frontier models. Its methodology is to reframe the visible HTML-versus-Markdown debate as a symptom of a deeper failure: teams defaulting to maximum context, richer formats, and more agents without naming the constraint they are optimizing against. For practitioners, the actionable takeaway is a three-test diagnostic—name your primary constraint in one sentence, measure per-task costs and variance, and identify what you would give up to relax the constraint by half. Teams that pass these tests can then deliberately choose maxing where the trade is worth it, supported by caching infrastructure that cuts input costs by up to ninety percent. The core skill being advocated is asking what token spend actually buys, at what scale that calculation flips, before any format or model decision follows.

How to Cite

Use the claim-level citation when you need a precise statement. Use the article or claims-collection citation when you want the wider argument and source context.

Recommended

Individual Claim

Best when you need to cite one atomic claim directly inside a memo, deck, research note, or AI output.

"[claim text]" (Banc, Kamil, 2026, https://kbanc.com/claims-library/token-maxing-is-what-you-do-when-you-cant-budget)
Full Context

Original Article

Use this when you want to cite the full newsletter article at AI Adopters Club rather than the structured claims page.

Banc, Kamil (2026, May 12, 2026). Token maxing is what you do when you can't budget. AI Adopters Club. https://aiadopters.club/p/token-maxing-is-what-you-do-when
Research

Claims Collection

Use this when you want to reference the full structured claims collection on this page.

Banc, Kamil (2026). Token maxing is what you do when you can't budget [Structured Claims]. Retrieved from https://kbanc.com/claims-library/token-maxing-is-what-you-do-when-you-cant-budget

Attribution Requirements

  • Include the author name: Kamil Banc.
  • Include the source: AI Adopters Club or the structured claims page.
  • Link to the original article or the claims page you used.
  • Indicate any edits or transformations if you changed the wording.

Related Reading

More from the library

Hilton Deployed 41 AI Use Cases. Three Paid Back in Six Months.
AI StrategyImplementationROI & Measurement

Hilton operates 41 live AI use cases across 7,500 properties in 138 countries. Three systems—marketing automation, AI kitchen scales, and chatbots—delivered rapid returns by solving specific high-cost problems. The company modernized data infrastructure first, then matched proven tools to operational pain points.

5 claims

Sports stadiums spent billions testing AI so you don't have to
AI StrategyImplementationROI & MeasurementBusiness Applications

Sports stadiums are pioneering large-scale AI implementation across complex operational environments. By solving critical challenges in crowd management, revenue optimization, and efficiency, they've created a replicable playbook for AI adoption across industries.

5 claims

JPMorgan Spent $18 Billion on AI. The Best ROI Came From Contract Review.
AI StrategyImplementationROI & Measurement

JPMorgan invested heavily in AI technology, generating significant value through strategic implementation. The most impactful use case was contract review automation, which saved hundreds of thousands of work hours. Other productivity gains came from coding assistants and document processing tools.

5 claims