Skip to content
p0.
Metered AI creates real user anxiety: cost is coupled to every action, and tokens are an invisible, abstract unit. The calm pattern is cost transparency: a visible balance, a cost preview before each turn, an early running-low warning, and a graceful out that never walls someone mid-thought or loses their work. Hiding the meter makes it worse.
Agent UXToken-cost anxiety

Designing for the anxiety of metered AI

Every turn costs money now, and users feel it. They hesitate before a long prompt the way they hesitate before a taxi meter. Cost transparency has quietly become a UX requirement, and almost nobody has designed it well.

  • Published 18 June 2026
  • Observed convention
  • Live demo
(01) The problem

For most of software history, using a feature was free at the point of use. You had paid for the licence or the subscription, and clicking a button cost nothing more. The meter, if there was one, ran in the background and someone else watched it.

Agents broke that. Tokens couple cost directly to action. Every message, every long document, every tool call spends something, and the user is the one watching the number. That is a genuinely new psychological condition, and it produces a measurable behaviour: people send shorter prompts, avoid ambitious requests, and abandon sessions, not because the product is bad, but because they are afraid of the meter.

The instinct is to hide the meter so people relax. It does the opposite. Uncertainty is the anxiety. The fix is honest information, paced well, with a soft landing, so the user always knows where they stand and never gets walled mid-thought.

(02) Live demo

Spend a few turns, both ways

Press send a few times in each mode. One buries the cost and walls you at zero. The other keeps you informed and lands you softly.

Tokens: 4200
Sure, here is a draft of that proposal section...

A number you cannot read, no sense of what a turn costs, and a hard wall that drops mid-thought. Every send feels like a gamble. That is the anxiety.

(03) The rules

What calm metering looks like

Six rules for showing cost without manufacturing dread.

Make the balance visible and legibleA bar beats a buried number. The user should know roughly how much runway they have without doing arithmetic or opening a billing page.
Preview the cost before they commitA small estimate next to the send button (this turn is about 600 tokens) lets people decide with information, instead of bracing after the fact.
Warn early, while there is still room to actSurface running low at roughly a quarter remaining, not at zero. A warning with no runway left is just an error message.
Never wall someone mid-thoughtA hard stop in the middle of a task destroys both the work and the trust. Degrade gracefully: save state, offer a top-up, let them continue the moment they do.
Translate tokens into something humanTokens are an abstraction. Where you can, map them to turns, messages, pages or pounds, the unit the user actually thinks in.
Make topping up frictionless and reversibleIf the calm path is to top up, that path has to be one tap and obviously safe. Cost anxiety is half about money and half about losing control.
(04) Questions

Frequently asked questions

Why do tokens make users anxious?
Because cost is now coupled to every action, and the unit is invisible and abstract. A user cannot see a token, cannot easily predict how many a request will burn, and often discovers the cost only after it is spent. That uncertainty produces the same hesitation as a taxi meter: people self-censor, send shorter prompts, or avoid the product entirely. The anxiety is a UX problem, not a pricing problem.
Should I just hide the token count to reduce anxiety?
No. Hiding the meter trades one anxiety (watching it drain) for a worse one (no idea where you stand, then a surprise wall). The calmer pattern is transparency with a soft landing: a visible balance, a cost preview, an early warning, and a graceful out that never loses work. Opacity erodes trust; honest, well-paced information builds it.
How do I show cost when I do not know it until the model responds?
You can estimate the input cost precisely before sending, and predict a typical output range from the task type and your own history. Show a range or a typical figure rather than false precision, then reconcile to the actual once the response lands. A good estimate before the action plus an accurate number after is far better than silence.
Is token-cost UX actually a recognised pattern?
It is widely practiced but barely documented. The trust and transparency spine of agent UX is peer-reviewed, but cost transparency specifically sits in near-open white space: most products handle it ad hoc and there is little authoritative guidance. That makes it both a real user need and an opportunity to define the pattern well rather than copy a bad default.
Does showing cost reduce usage and therefore revenue?
In practice the opposite tends to hold for trust-led products. Hidden meters and surprise walls cause churn and disputed charges; visible, fair metering with easy top-ups builds the confidence to use the product more. The goal is not to scare people off expensive actions, it is to remove the fear of the unknown that suppresses all actions.
Take it with you

Don’t just read this. Put it to work.

The whole series is distilled into one Markdown file: every pattern, the do and don’t rules, and how well each is evidenced. Download it into your project, or paste the link into any chat with your agent and tell it to improve your agent UX. It’s free, no sign-up, no attribution required.

Paste this into your agent

Use these Agent UX principles to review and improve our agent's interface: https://p0stman.com/agent-ux/agent-ux-principles.md

Part of the Agent UX series

We build the metered try-register-pay funnel into our own agent ventures, balance UX and all. If you are shipping a token-based product and the meter is scaring people off, that is a problem we have solved before.

Let's build

Metered products that people trust enough to use

Talk to PaulAbout your agent

Or write to hello@p0stman.com

AGENT INTERFACE ACTIVE · MCP: p0stman.com/api/mcp · 5 TOOLS REGISTERED · [DISCOVERY] llms.txt · agents.md · context.md · sitemap.xml · robots.txt · TavilyBot ALLOWED · ClaudeBot ALLOWED · GPTBot ALLOWED · PerplexityBot ALLOWED · [COMPREHENSION] JSON-LD schema · /api/ai/context · /api/ai/services · /api/ai/portfolio · [ACTION] book_discovery_call · submit_inquiry · get_services · get_portfolio · search_content · [A2A] AgentCard: /.well-known/agent.json · Task endpoint: /api/agent · A2A JSON-RPC 2.0 · navigator.modelContext REGISTERED · WebMCP: 5 TOOLS · INDEXNOW: 145 URLs · Bing NOTIFIED · [MANAGED AGENTS] Lead Researcher · AgentReady Auditor · SEO Writer · Weekly Reporter · Claude Sonnet 4.6 · Cloud containers · Outcome-based grading · Multi-agent orchestration · AGENT INTERFACE ACTIVE · MCP: p0stman.com/api/mcp · 5 TOOLS REGISTERED · [DISCOVERY] llms.txt · agents.md · context.md · sitemap.xml · robots.txt · TavilyBot ALLOWED · ClaudeBot ALLOWED · GPTBot ALLOWED · PerplexityBot ALLOWED · [COMPREHENSION] JSON-LD schema · /api/ai/context · /api/ai/services · /api/ai/portfolio · [ACTION] book_discovery_call · submit_inquiry · get_services · get_portfolio · search_content · [A2A] AgentCard: /.well-known/agent.json · Task endpoint: /api/agent · A2A JSON-RPC 2.0 · navigator.modelContext REGISTERED · WebMCP: 5 TOOLS · INDEXNOW: 145 URLs · Bing NOTIFIED · [MANAGED AGENTS] Lead Researcher · AgentReady Auditor · SEO Writer · Weekly Reporter · Claude Sonnet 4.6 · Cloud containers · Outcome-based grading · Multi-agent orchestration ·