The Context Problem

Token Economics and the Word That Means Everything

By Jessica Talisman, MLS · March 11, 2026 · Intentional Arrangement

An analysis of how AI token economics has turned context into a billing unit, the grammatical sprawl of the word across noun/verb/adjective roles, and why capacity is not coherence.

Core Concepts

Key Ideas

The article examines how a single word — context — has been stretched across three grammatical roles, each with distinct failure modes and cost implications.

Context-as-Noun

Context as the model's inputs and outputs — a long list where the AI model keeps a working memory for the conversation. The dominant usage in the industry. Cursor defines it as your model's inputs and outputs.

Context-as-Verb

The act of "contextualizing" — prepending background text to a prompt. Treats providing information as equivalent to establishing meaning. The limitation is that pasting documents is not the same as structured knowledge representation.

Context-as-Adjective

Phrases like "contextual AI" and "context-aware systems" that claim relational intelligence while the underlying noun may only deliver statistical proximity.

Context Rot

Documented by Chroma's 2025 study across 18 LLMs: as input length grows, model performance degrades even on simple tasks. Counterintuitively, shuffled context (randomly reordered sentences) produced better retrieval than logically organized context.

Neurosymbolic AI

The combination of data-driven and knowledge-driven techniques. Knowledge graphs and ontologies form the symbolic backbone; LLMs form the neural component. GraphRAG documented an 80% decrease in token usage vs. vector RAG.

Market for Lemons

Proposed by George Akerlof (1970). Applied to AI: foundation companies define the token, control coherence mechanisms, and bear no disclosure obligation. Buyers cannot inspect semantic quality before purchase.

Credence Good

Formalized by Darby and Karni (1973). AI processing is a credence good — quality cannot be verified even after consumption. One can evaluate output quality but cannot verify coherent vs. stochastic inference.

Pricing Comparison

AI Model Context Pricing

The spread encodes a market stance: larger context commands higher value, and the most capable reasoning earns a premium rate per token.

ModelOrganizationContext WindowInput $/MOutput $/MLicense
GPT-5.4 ProOpenAI1,050,000$5.00$180.00Proprietary
Claude Opus 4.6Anthropic1,000,000$5.00$25.00Proprietary
Gemini 2.5 FlashGoogle1,000,000$0.50$2.50Proprietary
Grok 4.1 FastxAI2,000,000$0.20$0.50Proprietary
Llama 4 ScoutMeta10,000,000$0.08$0.30Open-weight
Llama 4 MaverickMeta1,000,000$0.15$0.60Community
Grok 4xAI256,000$3.00$15.00Proprietary
FAQ

Frequently Asked Questions

Key questions derived from the article's analysis of context, token economics, and AI reliability.

What is context in information science?
Context, in information science, describes the relational structure that holds meaning in place. It is not a container but the fabric that connects one element to another. From its Latin root contextus — the act of joining together (con- + texere, to weave) — context encodes a relational theory of meaning.
How has AI token economics changed the meaning of context?
AI token economics has turned context into a billing unit. Foundation AI companies leveraged the context craze, pushing the onus of poor AI performance back on users. The cost of using AI is associated with tokens, and tokens are directly attached to context.
What are the three grammatical roles of context in AI discourse?
Context-as-noun: the model's inputs and outputs, a working memory list. Context-as-verb: the act of prepending background text to a prompt. Context-as-adjective: phrases like "contextual AI" and "context-aware systems" that claim relational intelligence while the underlying noun may only deliver statistical proximity.
What is context rot?
Context rot is the phenomenon where model performance degrades as input length grows, even on deliberately simple and controlled tasks. Chroma's 2025 study evaluated 18 LLMs and documented that every model tested showed nonuniform performance degradation as tokens accumulated.
What is context engineering?
Context engineering is the art and science of curating what will go into the limited context window from that constantly evolving universe of possible information. The guiding principle is finding the smallest possible set of high-signal tokens that maximize the likelihood of some desired outcome.
How does the "Market for Lemons" apply to AI context pricing?
George Akerlof's 1970 theory describes markets where sellers know more about quality than buyers. In AI context pricing, foundation AI companies define the token, control the mechanisms that determine whether tokens cohere, and bear no disclosure obligation regarding coherence.
What is a credence good in the context of AI?
A credence good, formalized by Darby and Karni (1973), is a good whose quality cannot be verified even after consumption. AI token processing is a credence good: one can evaluate output quality after the fact, but cannot verify whether that output reflects coherent processing or stochastic inference.
How do knowledge graphs reduce token costs?
Passing a well-formed knowledge graph subgraph into a context window delivers more inferential signal per token than passing equivalent raw text. Research documented an 80% decrease in token usage when graph-based retrieval replaced conventional vector RAG methods.
Why is capacity not the same as coherence?
The market has priced capacity — how many tokens the model can see at once. What it has not priced is whether the relational structure within those tokens is coherent, navigable, or epistemically sound. Llama 4 Scout offers 10M tokens but effective context degrades at ~32,000 — a 300x gap.
What did Gartner recommend for AI value derivation?
Gartner identified three pillars: (1) A clear AI ambition tied to return on intelligence. (2) Trusted data systems and governance as the foundation for return on integrity. (3) People equipped with the skills to execute. Gartner also named composite and neurosymbolic AI as one of three future AI scenarios.
Glossary

Defined Terms

Key terms from the article's analysis of context, token economics, and AI reliability.

Context
The relational structure that holds meaning in place. From Latin contextus: the act of joining together.
Token
The smallest unit of information an AI model processes — a snippet of text from a single character to an entire word.
Context Window
The amount of text, in tokens, that a model can consider at any one time. IBM: "the equivalent of its working memory."
Context Rot
Performance degradation as input token length grows, documented by Chroma across 18 LLMs.
Context Engineering
The art and science of curating the optimal set of tokens for the context window.
Neurosymbolic AI
Combination of data-driven (neural) and knowledge-driven (symbolic) techniques.
Knowledge Graph
A structured knowledge representation using explicit relationships. More inferential signal per token than raw text.
Market for Lemons
Akerlof's 1970 theory of markets with asymmetric information quality. Applied to AI context pricing.
Credence Good
A good whose quality cannot be verified even after consumption (Darby & Karni, 1973).
Information Asymmetry
Sellers possess more information about product quality than buyers. Foundation companies control coherence mechanisms.
Context-as-Noun
Context as the model's inputs and outputs — a working memory list. The dominant usage.
Context-as-Verb
The act of contextualizing — prepending background text to a prompt.
Context-as-Adjective
Phrases like "contextual AI" that claim relational intelligence while delivering statistical proximity.
Byte Pair Encoding
A tokenization method that merges frequent character pairs to form tokens.
Compaction
Summarizes conversation history and reinitializes the context at the cost of some fidelity.
How-To

Context Engineering & Neurosymbolic Solutions

Practical steps for managing context costs and building coherent AI systems.

Identify the grammatical role of context in your use case

Determine whether you need context-as-noun (working memory for conversation), context-as-verb (prepending background to prompts), or context-as-adjective (context-aware system design). Each role has distinct failure modes and cost implications.

Apply context engineering techniques to manage token costs

Use compaction to summarize and reinitialize context. Implement structured note-taking to persist critical state outside the window. Deploy sub-agent architectures to isolate focused tasks in clean context windows.

Build neurosymbolic knowledge structures for coherence

Pass well-formed knowledge graph subgraphs into context windows to deliver more inferential signal per token than raw text. Use ontologies derived from relational databases to reduce overall LLM usage costs.

SPARQL Queries

Explore the Knowledge Graph

Pre-built SPARQL queries to explore the article's knowledge graph. Click "Run live query" to execute against URIBurner.

▼ Describe the Article
Run live query 
PREFIX schema: <http://schema.org/>

DESCRIBE <https://open.substack.com/pub/jessicatalisman/p/the-context-problem#article>
        <https://open.substack.com/pub/jessicatalisman/p/the-context-problem#jessicaTalisman>
▼ Construct the FAQ Subgraph
Run live query 
PREFIX schema: <http://schema.org/>

CONSTRUCT {
  ?q a schema:Question ;
    schema:name ?name ;
    schema:acceptedAnswer ?a .
  ?a a schema:Answer ;
    schema:text ?text .
}
FROM <https://linkeddata.uriburner.com/DAV/demos/daas/the-context-problem-jessica-talisman-substack.ttl>
WHERE {
  ?q a schema:Question ;
    schema:name ?name ;
    schema:acceptedAnswer ?a .
  ?a schema:text ?text .
}
▼ AI Models by Organization
Run live query 
PREFIX schema: <http://schema.org/>

SELECT ?model ?modelName ?orgName ?contextWindow ?inputPrice ?outputPrice
WHERE {
  GRAPH <https://linkeddata.uriburner.com/DAV/demos/daas/the-context-problem-jessica-talisman-substack.ttl> {
    ?model a schema:SoftwareApplication ;
      schema:name ?modelName ;
      schema:manufacturer ?org .
    ?org schema:name ?orgName .
    OPTIONAL { ?model :hasContextWindow ?contextWindow }
    OPTIONAL { ?model :hasInputPricing ?inputPrice }
    OPTIONAL { ?model :hasOutputPricing ?outputPrice }
  }
}
ORDER BY ?orgName ?modelName
Knowledge Graph

KG Explorer

RDF Graph Workbench

0 nodes / 0 links
Person
Organization
Concept
Article
Software
⚙

Explore Knowledge Graph using SPARQL

Choose a named graph and query recipe, edit the SPARQL if needed, then open the encoded URIBurner query.

Run live query
SPARQL Editable query
SELECT queries use text/x-html+tr. DESCRIBE and CONSTRUCT queries use text/x-html-nice-turtle, matching the SPARQL format guidance in the skill contract.
Tools Used

Tools & Technologies

D3.js v7

Data-driven document manipulation for interactive graph visualization

URIBurner

Linked Data resolver and SPARQL endpoint for knowledge graph exploration

OpenLink Virtuoso

Universal server for RDF, SQL, and SPARQL data management

kg-generator skill

Knowledge Graph generation from web content using schema.org terms

rdf-infographic-skill

Interactive HTML infographic generation from RDF knowledge graphs

RDF / SPARQL

W3C standards for knowledge representation and query