EDITION 2026 · H2
AI TOKEN BURN INDEX 2026
The public ranking of how money actually leaves your account when you use AI. Updated by edition, argued about constantly.
METHODOLOGY: editorial scores based on public per-token pricing and observed usage patterns. This is commentary with arithmetic, not telemetry. If your habit isn't here, it's not innocence — it's a backlog.
MOST EXPENSIVE AI HABITS
Re-pasting the entire conversation "for context"
The model already had the context. Now it has it twice, at input prices.
Agents left on auto-approve overnight
You woke up to 340 tool calls and a refactor nobody asked for.
A 2,000-token system prompt for a yes/no question
The answer was 'no'. The invoice was not.
"Rewrite the whole file" instead of asking for a diff
Output tokens cost 4-8x input tokens. Rewrites are output.
Transcript → summary → summary of the summary
Recursive compression, expansive billing.
Regenerating 14 times instead of fixing the prompt once
Slot-machine prompting. The house always wins.
Politeness rituals: please, thank you, sorry to bother you
Courtesy is free between humans. This is not between humans.
Decorative markdown headers in internal prompts
The model does not care about your emoji dividers. Your CFO might.
PROMPT OBESITY SCORE — BY ROLE
Brand voice documents pasted into every single request.
Vision statements where instructions should be.
Frameworks. So many frameworks.
User stories about the prompt inside the prompt.
Lean prompts, then pastes the entire monorepo.
Efficient. Suspiciously quiet about their notebook token usage.
AGENT BURN RISK
Multi-agent crew with a 'reflection' step
Five agents discussing. Conclusion: 'it depends'.
Coding agent in a retry loop on a failing test
Attempt #23 looks a lot like attempt #4.
Research agent with unbounded web browsing
It read the whole internet. It cited two tweets.
Agent re-reading the full repo every task
Context is not RAM. It is billed like a hotel minibar.
Scheduled agent that mostly reports 'no changes'
Paying daily to be told nothing happened.
Single agent, scoped task, capped iterations
This is what discipline looks like. Boring, isn't it?
WORKFLOWS THAT BURN MORE THAN EXPECTED
| Workflow | The expectation | The reality |
|---|---|---|
| Daily 'summarize all Slack channels' | A few cents | Every message re-sent as input, every day. Compounding. |
| AI meeting notes for every meeting | Cheaper than a scribe | The summary of the standup cost more than the standup. |
| Agent code review on every commit | Quality gate | Full-diff context per commit x commits per day x output verbosity. |
| Generate 50 variants, pick 1 | Creative exploration | 49 deleted drafts, billed at output rates. |
| Chatbot with full history in every turn | It just remembers | Nothing 'just' remembers. Input tokens remember, per turn. |