Failure mode
Raw pages into context
A web fetch returns the whole page when the agent asked for one fact; raw HTML can be tens of thousands of tokens per page.
Token Economics
Know what AI tokens buy you.
Token economics is the discipline of knowing what AI tokens buy you: how compute is priced into tokens, how organizations allocate them, how agents consume them, and how that spend is measured against outcomes.
The Efficient Token Tree
RootYour token bill
Context
Failure mode
Full tool-output dumps
Ten thousand rows enter context when five were needed.
Failure mode
Every tool schema, every turn
Classic MCP re-sends the full tool catalog each turn whether or not it's used.
Cache
Failure mode
Prefix edits mid-session
Adding one tool or editing the system prompt re-bills the whole cached prefix at full price, with no warning.
Failure mode
Token cuts that bust the cache
Compression aimed at the prefix trades a ~90%-discounted token for a full-price one: measured at minus 38% tokens, plus 6.8% cost.
Model choice
Failure mode
Frontier model for everything
Reading files, running tests, and formatting docs at frontier prices.
Failure mode
Batch escalation
Whole workloads bumped to the expensive tier because one item was hard.
Harness
Failure mode
Unmeasured fixed payload
Tens of thousands of scaffolding tokens paid every session before the first prompt (the 32,800 vs 6,900 gap).
Failure mode
Compaction churn
One harness compaction retune roughly doubled a real bill via file re-read amplification (Codex v0.118 regression).
Failure mode
Unverified plugin claims
Token-saver plugins advertising 60–90% cuts; the one independent quality-controlled A/B test measured one of them net cost-negative at low reasoning effort.
Governance
Failure mode
Unbounded loops and retries
A documented retry storm consumed $1,434 in one day; harness retry ceilings have been raised, not lowered.
Failure mode
A bill nobody can explain
Token counts that don't reconcile to the invoice; service tiers that double price invisibly (three vendors now sell a 2x latency tier).
Failure mode
Cutting before understanding
Optimizing whatever is most visible.
The numbers behind the tree
~66x
input-token spread, same model, leaner harness scored higher
Terminal-Bench, 2026 ↗4.7xClaude Code ~32,800 vs OpenCode ~6,900 scaffolding tokens
Systima, 2026 ↗≈80%of a real provider bill came from cache traffic
arXiv:2607.12161, July 2026 ↗−38% / +6.8%tool-output tokens removed, cost increased after the cached prefix broke
arXiv:2607.12161, July 2026 ↗≈0.1xbase input price for cache reads on both major providers
Anthropic and OpenAI pricing, 2026-08 ↗≈2xbill after a Codex CLI compaction retune
Codex CLI v0.118 regression, 2026 ↗$1,434/dayconsumed by a documented retry storm
arXiv:2604.16646, 2026 ↗$1,500/moper-tool cap at Uber after about four months of budget burn
TechCrunch, 2026 ↗+7.6%cost at low effort in an A/B test of a plugin claiming 60–90% cuts
JetBrains, July 2026 ↗The field guide keeps growing
Browse practicesLonger reads on the blog
HN I gave Opus 5.5 one prompt and six hours to visualize Invisible Cities
Claude Opus 5.5 vs GPT-6 Astra one-shot Invisible Cities by Italo Calvino in Three.js: AI design, UX, data visualization and explorable explanations.
Your agent session transcripts are precious, keep them
AI vendors collect every coding session they can, yet most teams discard their own. Quesma Shipper collects them into Amazon S3.
Token economics in Amsterdam: Inside the first Tokenomicon
The Linux Foundation expands its FinOps work with a Tokenomics organization focused on AI spend. It held its first in-person conference in Amsterdam.