Token & time consciousness · A Claude Code skill

Before you run it,
ask: at what cost?

An AI agent task is cheap to start and easy to under-think. A one-line request can quietly become hundreds of thousands of tokens — and real electricity, carbon, and water.

At What Cost is a gate-check skill: before a major LLM task, it estimates the toll — tokens, dollar cost, wall-clock time, CO₂, and water — as honest ranges, so accepting that toll becomes a deliberate choice instead of a default. Try the estimator below.

Try the estimator ↓ View the skill on GitHub
Interactive

Estimate a task's footprint

Pick a task shape and a model. The skill carries low / expected / high ranges through every metric — never a single false-precision number.

Pricing from Anthropic's published per-million-token rates.
MetricLowExpectedHighConfidence
Roughly equivalent to (expected case — approximate)
Read this: pre-task token & time estimates are directional (±2–5×), not ±10–20% — an agent discovers its own scope as it runs. CO₂ & water are order-of-magnitude. Cost is the firmest number. Scope is this task's inference only — not model training, your device, or the network.
The idea

How the gate-check works

It runs only when you ask for it — explicitly (/at-what-cost) or implicitly ("footprint check before we start"). It informs a decision; it never blocks.

01

Classify

Match the task to a shape — quick lookup, single-file edit, multi-file refactor, research report, multi-agent workflow.

02

Estimate

Anchor the input on what's already in scope; take the output side from the class band. Carry low / expected / high throughout.

03

Compute

Cost from published rates, then energy → CO₂ → water from cited coefficients. Every metric is a range.

04

Decide

See the table, accept the toll, rescope, or cancel — a conscious choice before any tokens are spent.

05

Reconcile

After the run, the skill reads the actual token usage, compares it to the estimate, and logs it — calibrating future estimates over time.

The honest part

It's an awareness tool, not a calculator

Some of these numbers genuinely cannot be precise — and pretending otherwise would be the dishonest move. Here's exactly how much to trust each one.

MetricRealistic accuracyWhy
Cost ($)firm given tokensPublished rates × tokens — the rates are exact; error is whatever the token estimate's error is.
Tokens / time±2–5× (directional)Agentic scope is discovered at runtime — reads, tool calls, retries, and sub-agents aren't knowable up front.
CO₂eq / waterorder-of-magnitudeNo provider publishes per-token energy or water; coefficients are third-party, dated, and region-sensitive.
This is why the skill can't promise a ±10–20% figure on tokens, time, CO₂, or water — and says so in every output. What it can do is keep you in the right order of magnitude, lead with the firm number, and get better the more you use it.

Coefficients are cited & tunable: Anthropic pricing (per-MTok), 2025 LLM inference-energy studies (e.g. arXiv:2504.17674), IEA/Ember grid carbon intensity, and Ren et al. “Making AI Less Thirsty” (water). Each value in factors.json carries its own source and date.