‹ Back to Blog

What Did the Cockpit Cost?

Claude's own estimate of the energy, carbon and water behind Junquer's main menu

junqueraidev-log

Yesterday I posted Is This AI Slop?, about the two days I spent directing Claude through Junquer’s new main menu. I shared it on LinkedIn and got some thoughtful questions back. One comment asked something I had considered, but had no real answer for. What did the menu really cost? Not what I paid for tokens, but the whole cost behind the scenes, environmental included, and how that compares with a human artist making the same thing.

I spent a very small amount of time researching it on my own to provide a response, but then I got curious about what Claude itself had to say about the matter. So I asked it. Claude Code keeps a transcript of every session, down to how many tokens each call read and wrote, so it can count its own work. I sent Opus 5.5 one prompt, and about fifteen minutes later it handed back the report below.

Read this skeptically

I have not checked most of it. I clicked through the links to check they led to the cited sources, but did not verify that the numbers it claims from those sources are in any way accurate. Its token counts are in the range I expected (based on my own tracking). I haven’t verified the energy per token estimates. The report itself says its main numbers are uncertain by a factor of four or five either way. It also points out that Anthropic publishes no energy, carbon or water figures, so everything on the Claude side is inferred from other companies’ models and hardware.

My gut reaction to the time it would take a human to do the same work is that Claude’s numbers are on the high side, especially for a highly qualified human. But I also don’t think I could have hired a human to produce the menu work and have it done in two days. I would have guessed in the one to two week range.

My goal was not to create a credible source of information for this discussion. I wanted to see what the model would do, what sources it would cite, what assumptions it would make, and ultimately what numbers it would present in its analysis. I’m posting this because I think the way it answered is interesting and valuable, whether or not its answer is at all accurate. Treat it as a place to start a conversation.

Everything below the prompt is exactly what Claude wrote. I haven’t changed a word. It made all of the figures itself. A couple of lines point at files in Junquer’s repository, which isn’t public.

It also only answers half of the question. It covers energy, carbon and water. It doesn’t get into what I’d pay a human artist, where that money goes afterwards, or the comment’s other good idea: commission the first sketch from an artist, then refine it with AI from there. Those are equally, if not more, important considerations that I address in the LinkedIn comment thread. But I wanted to keep this post focused on exactly what Opus said.

If you know this area and something here looks wrong, please say so in the comments below or on Discord. I’d like to know which numbers hold up and which don’t.

The prompt

give me a report on the estimated environmental costs of all of the updates made to the main menu in the last few days. and compare them to the estimated human environmental costs to produce the same. include all assumptions made in the estimations, and sources for data used

The report

Junquer · main menu · Oct 20 to 21, 2026 · estimate

What the cockpit cost to make

The cockpit main menu took about 1,800 Claude calls over two days. My best estimate is 10 kWh of electricity, 4 kg of CO2e and 38 litres of water, each uncertain by a factor of four or five either way. A person doing the same work from home would use somewhat more. A person commuting to an office would use about a hundred times more.

62 menu commits1,814 Claude calls268 M tokens handledhuman version about 180 hours
10 kWhelectricity, range 2 to 41
4 kg CO2eabout 10 miles in an average car18
38 Lwater, range 6 to 250

The short answer

Here are the four versions side by side. Each human row is the same piece of work, which I put at about 180 hours for one skilled person (range 111 to 260). They differ only in what gets counted.

Who made itEnergyCO2eWater
Claude, as it was made10 kWh4 kg38 L
A person at home, computer only27 kWh9 kg85 L
A person in an office, driving in620 kWh + heat500 kg2,000 L
A person's share of a US lifen/a360 kg58,000 L

The range charts further down show how wide each of these is. The Claude row and the at-home row overlap, so I would not claim the AI route was cleaner than a remote worker. It was no worse, and it took about 26 hours where a person would have needed a month.

What I counted

I read the token counts straight out of the Claude Code transcripts on this Mac. Every assistant call records how many tokens it read fresh, how many it read back from the prompt cache, and how many it wrote. I took every session that changed the main menu between Oct 20 and the end of Oct 21, including its subagents and the workflow that rebuilt the window wall. Where a session wandered onto other work, I cut it at the timestamp where the work changed.

WorkCallsWrittenRe-readkWh
Cockpit: renders, live Earth, wear, downlink, lamps, sound1,204802 k183 M4.8
Ship and station models from the store packs202131 k29 M0.8
Income line on the conn14190 k20 M0.5
Directorate money mark8451 k9 M0.2
Rival live view, built and not used7129 k12 M0.3
RESEARCH typed at the conn5720 k5 M0.1
Two desk-shot fixes5529 k4 M0.1
Menu total1,8141.15 M262 M6.8

The kWh column uses the middle per-token rates from the next section, before the cross-check that lifts the headline to 10. I left out two stretches that sat inside menu sessions but were other work: the in-game HUD restyle (0.6 kWh) and the claim depletion economy (1.0 kWh). Yesterday's write-up of the menu's evolution came to 0.2 kWh. It is not counted either, because it is not an update to the menu.

Three smaller costs are in the headline. Your Mac ran Blender renders, tests and headless screenshots for about five and a half hours of tool time, which I put at 0.3 kWh. The three Codex reviews used about 39,000 tokens each, roughly 0.04 kWh in all. Your own 26 hours of directing it are not counted on either side, because you would have directed a human artist too.

Where the AI energy goes

An agent re-reads its whole conversation on every call. The average menu call here re-read 145,000 tokens from the prompt cache, read 2,500 new ones, and wrote 635. So 98% of the tokens handled were re-reads of context it had already seen.

Tokens handled, 268 million re-read 262 M (98%) new 4.6 M, written 1.2 M Energy, middle case by token, 6.8 kWh re-read 4.2 kWh (61%) new 0.9 written 1.7 (25%) re-read from cache new input written output
Re-reading context is almost all of the tokens. Each re-read token is cheap, so it comes to about 60% of the energy. Writing is a sliver of the tokens and a quarter of the energy.

That re-read share is the least certain number in the report. No one has measured what a cached token costs in energy compared with a fresh one.7 Anthropic charges 5% of the fresh price for a cache hit on this model,6 but a price is not a meter. I used 8% of the fresh-token energy as the middle case, taken from the one public attempt at a Claude Code footprint,8 with 3% and 15% as the ends. Long sessions push this term up, because every call carries the whole conversation so far. The cockpit session ran out of room and restarted from a summary many times over its two days.

Two ways to estimate it

By token

I took per-token rates from the published per-query studies. Jegham and colleagues measured Claude 3.7 Sonnet at 0.95 Wh for a short query and 5.7 Wh for a 10,000-token prompt with a 1,500-token answer.3 A straight-line fit through their three Claude points gives about 2.7 mWh per written token and 0.15 mWh per read token. Google's own measured median for a Gemini prompt is far lower, at 0.24 Wh,1 and Epoch puts a 100,000-token input near 40 Wh.2 A Microsoft study of frontier models found a median of 0.31 Wh, rising to 3.9 Wh for queries 15 times longer.4 My rates span those.

Per tokenLowMiddleHigh
Written0.5 mWh1.5 mWh3.0 mWh
Read fresh0.1 mWh0.2 mWh0.4 mWh
Re-read from cache0.003 mWh0.016 mWh0.06 mWh
Menu total1.8 kWh6.8 kWh21 kWh

By dollar

At list prices the menu's tokens would have cost $111. Suppose 40 to 60% of that price pays for compute, and each compute dollar buys 0.25 to 0.6 kWh at the wall, which is roughly what renting a datacentre GPU costs against what it draws. That gives 11 to 40 kWh, with 22 in the middle.

The dollar method runs about three times higher than the token method. It ignores Anthropic's cheaper TPU and Trainium hardware, and the token studies assume very low GPU use, so each has a reason to be off in its own direction. I took 10 kWh as the headline, between the two middles. For the range I kept the lowest and highest ends of both methods, 2 to 41 kWh.

The human version

To compare, I priced the same deliverables as hours for one person who can model and render in Blender, write three.js and TypeScript, and make sound. In practice that is two or three specialists. These are my estimates, not measured times.

Piece of workHours
First cockpit in CSS, Earth in the window4 to 8
Rendered window wall and boom arm16 to 36
Rendered console housing, its dents and worn edges10 to 24
Live polluted Earth, stars, elevators, junk, ships24 to 56
Ship and station models converted and placed6 to 14
Wear passes, C + L heart, scuffs, rag box, bar lamps14 to 34
Downlink both ways, LINK lamp, console print strip12 to 26
Orbit hum, failing tube, hull knocks6 to 16
Credit line, income rate, money mark font8 to 18
RESEARCH command, shot fixes, texture versioning5 to 12
Tests and review6 to 16
Total111 to 260

The rival live view is left off the human side. A person would not build the Earth twice to pick one.

I then counted that time three ways.

At home
Only the computer. A desktop doing 3D work with its monitor draws about 150 W in the middle case, a laptop 75 W, a render box 250 W.13 Home heating and lighting are left out, since the person is home either way.
Office and car
An average US office uses about 6,900 kWh of electricity and 2,800 kWh of other fuel per worker per year.17 I spread that over 2,000 working hours. Add a 13.4-mile commute each way19 in a car that emits 400 g of CO2 a mile,18 for 22 working days.
Share of a life
The method from Tomlinson and colleagues.13 Take a US resident's whole yearly footprint, 17.5 t CO2e15 and 2,842 m³ of water,16 and charge the work its share of the year's 8,760 hours.

Side by side

Claude 4 kg Person at home 9 kg Share of a life 360 kg Office and car 500 kg 0.11101001,000 kg
Carbon, log scale. Claude's range overlaps the person working at home. The office and share-of-life versions sit about a hundred times higher.
Claude 38 L Person at home 85 L Office power 2,000 L Share of a life 58,000 L 1101001k10k100k L
Water, log scale. Datacentre cooling plus the water power plants use comes to tens of litres. Charging the work its share of a person's food and goods puts it in the tens of thousands.

Which comparison is fair

I think the at-home row is the fair one, and on that row the two routes come out about even. The share-of-life method gives the most dramatic ratio, and it is the one in the much-quoted Scientific Reports paper.13 But a person eats and heats their home whether or not they model a cockpit. Sasha Luccioni made the same objection: you cannot charge a person's whole lifetime footprint to their job.14 The office and car costs are real, but a person can avoid them by working from home.

The honest counterfactual is probably neither. You would not have hired a month of artist time for a main menu. Without Claude the menu would have stayed the amber terminal. Seen that way the 4 kg is pure addition, about as much as driving ten miles.

The Claude numbers are also the shakier ones. Anthropic runs on AWS and Google Cloud20 but publishes no energy, carbon or water figures,22 so every number on the Claude side is inferred from other companies' models and hardware.

Assumptions

AssumptionLow / middle / highBasis
Energy per written token0.5 / 1.5 / 3.0 mWhGoogle, Epoch, Jegham fits1 2 3
Energy per fresh input token0.1 / 0.2 / 0.4 mWhJegham fit, Epoch long input2 3
Cached token as share of fresh3% / 8% / 15%Secondary estimate; unmeasured7 8
Cache write counted as fresh inputsame rateMy assumption
Share of price that is compute40% / 50% / 60%My assumption
kWh per compute dollar0.25 / 0.4 / 0.6My assumption from GPU rental against draw
Headline datacentre energy1.8 / 10 / 40 kWhEnds of both methods, middle between them
Facility overheadincludedPer-query studies already include PUE1 3
Grid carbon0.287 / 0.348 / 0.45 kg/kWhJegham AWS, eGRID US, gas-heavy guess3 11
Hardware manufacture42 g/kWh addedGoogle's embodied share1
Water per datacentre kWh3.3 / 3.7 / 6.3 LAWS, Google on-site plus generation water1 3 9 12
Facility PUE in the sources1.09 to 1.14Google, AWS9 10
Mac and Codex0.2 / 0.3 / 0.75 kWh5.5 h tool time at 20 to 60 W; 3 reviews
Human hours111 / 180 / 260My estimate, table above
Workstation draw75 / 150 / 250 WTomlinson laptop and desktop figures13
Office energy per worker-hour3.45 kWh power, 1.4 kWh gasCBECS 2018 over 2,000 h17
Gas carbon0.181 kg/kWhEPA factor, 53.06 kg/MMBtu21
Commute13.4 mi each way, 400 g/miNHTS 2022, EPA18 19
US per-capita footprint14.2 t CO2 / 17.5 t CO2eOur World in Data, 202415
US per-capita water2,842 m³ a yearHoekstra and Mekonnen16

What is not counted

  • Training the model. Anthropic gives no figure. Mistral's lifecycle study folds training into 18 months of use, so it cannot be split per token.5 If training added as much again as inference, Claude's carbon would double to about 8 kg. That is still in the at-home person's range.
  • Networking, your phone and the desk frame. These are small, and present in both routes.
  • Your own hours. You gave 53 directions over about 26 hours. You would have directed a human artist too, and that would have taken weeks of reviews.
  • Rare-metal depletion. Mistral reports it, but nobody has a comparable human figure.5
  • Market-based carbon. With renewable contracts Google reports 94 g/kWh.1 That would cut Claude's carbon to about 1.4 kg. I used grid averages on both sides so the comparison stays like for like.

The script and the raw token counts are committed beside this report in docs/reports/assets/2026-10-22-menu-footprint/, so the numbers can be re-run with different rates.

Sources

  1. Google Cloud, "Measuring the environmental impact of AI inference", 21 Aug 2025, and Elsworth et al., arXiv:2508.15734. Median Gemini text prompt 0.24 Wh, 0.03 g CO2e, 0.26 mL; PUE 1.09; WUE 1.15 L/kWh; embodied included. blog, paper
  2. Josh You, Epoch AI, "How much energy does ChatGPT use?", 7 Feb 2025. About 0.3 Wh a query; about 2.5 Wh at 10k input tokens, almost 40 Wh at 100k. epoch.ai
  3. Jegham et al., "How Hungry is AI?", arXiv:2505.09598 v6, 24 Nov 2025. Claude 3.7 Sonnet 0.95, 2.99 and 5.67 Wh per query; AWS PUE 1.14, WUE 0.18 on-site and 5.11 off-site, 0.287 kg/kWh. arxiv.org
  4. Oviedo et al., arXiv:2509.20241, in Joule 2026. Frontier models median 0.31 Wh a query; 15x longer queries 3.91 Wh. arxiv.org
  5. Mistral AI, "Our contribution to a global environmental standard for AI", 22 Jul 2025. 1.14 g CO2e and 45 mL per 400-token answer; 20.4 kt CO2e for training plus 18 months. mistral.ai
  6. Anthropic, prompt caching documentation, read 22 Oct 2026. Cache hits on Opus 5.5 at 0.05x base input; 1-hour writes at 2x. platform.claude.com
  7. Tian et al., "Cache Your Prompt When It's Green", arXiv:2505.23970 v3, Apr 2026. Caching avoids recomputation but adds storage carbon; no per-token ratio. arxiv.org
  8. gwittebolle, "How much CO2 does a Claude Code session actually emit?", DEV Community. Secondary; uses cache reads at 0.08x input energy and says no study measures it. dev.to
  9. AWS Sustainability. 2025 PUE 1.14; WUE 0.12 L/kWh withdrawn. aws.amazon.com
  10. Google data centre efficiency. 2025 fleet PUE 1.09. datacenters.google
  11. US EPA, eGRID2023 summary data. US output rate 770.9 lb CO2e/MWh, about 0.348 kg/kWh. epa.gov
  12. Li et al., "Making AI Less Thirsty", arXiv:2304.03271 v5, 26 Mar 2025. US average water for electricity 3.14 L/kWh. arxiv.org
  13. Tomlinson, Black, Patterson, Torrance, "The carbon emissions of writing and illustrating are lower for AI than for humans", Scientific Reports 14:3732, 2024. 75 W laptop, 200 W desktop; 1.7 kg CO2e per US person-hour. nature.com
  14. VentureBeat, "How much energy does AI use compared to humans?", 22 Sep 2023, quoting Sasha Luccioni. venturebeat.com
  15. Our World in Data. US 2024 CO2 per capita 14.2 t; GHG including land use 17.5 t CO2e. CO2, GHG
  16. Hoekstra and Mekonnen, "The water footprint of humanity", PNAS 2012. US consumer 2,842 m³ a year. doi.org
  17. US EIA, CBECS 2018, tables C22 and B1. Offices 13.6 kWh/ft², 507 ft² per worker, 65.6 kBtu/ft² total. C22, B1
  18. US EPA, "Greenhouse gas emissions from a typical passenger vehicle". About 400 g CO2 a mile. epa.gov
  19. US DOT, NHTS 2022 Summary of Travel Trends, table 7-4. Average commute trip 13.43 miles. rosap.ntl.bts.gov
  20. SiliconANGLE, Anthropic's Google TPU deal, 23 Oct 2025. Anthropic's compute runs on AWS and Google Cloud. siliconangle.com
  21. US EPA, GHG Emission Factors Hub. Natural gas 53.06 kg CO2 per MMBtu. epa.gov
  22. Ketan Joshi, commentary, 13 May 2026. Notes Anthropic publishes no energy or emissions data. ketanjoshi.co

Comments

Loading comments...