Show your work

Every number on a receipt carries one of four labels. This page shows where each one comes from. It is meant to be boring.

The four labels

No mark: measured. Read from your logs. Exact counts.

≡ Priced: public API list price × measured tokens. What the same usage would cost at list price, not what you paid.

≈ Estimated: always a range, at most two significant figures, with the method below.

✶ Satire: made up on purpose. Always red, always labeled as made up.

Every receipt is stamped with the method version and the date of its prices.

Prices

List prices in USD per million tokens, as of September 24, 2026. They are checked against the providers' price pages before every release.

A model missing from the table is priced as the closest listed one and marked *. A model with no listed family says "not priced", never $0.

Electricity (method v1)

Providers publish no per-token energy figures for these models. We use per-token coefficients by token type, calibrated so that one published Claude Code log (3.2 billion tokens, 96% cache reads) reproduces its author's central estimate of 170 kWh. The low and high ends follow the author's 70–330 kWh range.

Low = central × 0.41. High = central × 1.94.

Cache reads are the largest uncertainty: bounds of 1% to 25% of a fresh token.

Sanity checks: about 50 kWh per billion agent tokens (20–100), and about 41 Wh for a median Claude Code session. A result far outside these means the parser is wrong, not the world.

Water and CO₂

Shown in the detail view only, never on the share card. Water for on-site cooling: kWh × 0.2–2 L/kWh. Off-site water is larger and disputed; it is shown as a separate note and never added in. CO₂e: kWh × 0.34–0.42 kg/kWh (US, location-based); market-based figures run lower.

Satire formulas

RAM-X, your share of the shortage: tokens ÷ 60 trillion × 0.95 × 0.42 dollars per stick. 60 trillion is Meta's reported 30-day token usage, 0.95 the Q1 2026 rise in DRAM contract prices, and 0.42 the Coefficient of Vibes. The formula is absurd on purpose. The price rise is real. That you caused it is not.

Under RAM-X, one more ✶ line from a rotating pool: a real, dated AI event and your share of it, tokens ÷ 60 trillion × 0.42 × the event's size. The event is real. Your share is not. A receipt joke and a news line rotate with it, and nothing repeats until you've seen the whole pool.

What building this site cost

Token Damage, this site included, was written with an AI coding agent. So we ran it on its own logs. Our receipt, as of September 24, 2026

Sources

Citations stay in their original language.

• Hausfather, "The real energy use of agentic AI", 5 Aug 2026 (1,138 prompts → 14,000+ calls, 3.2B tokens, 96% cache reads, 0.4% output, 170 kWh (70–330), ~150 Wh/prompt, ~600× a chat prompt)
• Couch, "Electricity use of AI coding agents", 20 Jan 2026 (median session 41 Wh, median day 1.3 kWh)
• Google, "Measuring the environmental impact of AI inference", 21 Aug 2025 (0.24 Wh, 0.26 mL, 0.03 gCO₂e)
• Altman, "The Gentle Singularity", Jun 2025 (0.34 Wh per query)
• Epoch AI, "How much energy does ChatGPT use?", Feb 2025 (~0.3 Wh)
• Mistral, environmental contribution, Jul 2025 (1.14 gCO₂e, 45 mL per 400-token response)
• LBNL, US Data Center Energy 2025 update, Jun 2026 (11.8% by 2030, 9.5–15.3%)
• DOE/LBNL 2024 report (4.4% in 2023; 6.7–12% by 2028)
• EcoLogits methodology (ISO 14044 LCA; output-token regression; ignores cache reads)
• Jegham et al. fit used by the dev.to analysis (1.35e-4 Wh/input token, 2.88e-3 Wh/output token)
• Off-site water (Ren, UC Riverside)
• TrendForce, 1 Jun 2026 (DRAM contract +93–98% QoQ in 1Q26; +58–63% expected 2Q26; industry revenue $97B; Samsung 38.5%)
• TrendForce, 2 Feb 2026 (90–95% forecast for 1Q26)
• Micron exits Crucial, 3 Dec 2025