UsesLevel 3 of 6 useUse, level 3 of 6Used to calculate 10 entries, level 4Calculated with 5 entries, levels 1 and 2Database structure 
6Territories
5Organisations
4Occupations and activities
2Machine operations
1Calculation factorsHave a coding agent work, 1 hour
One hour of a coding agent that chains 120 steps on its own and each time re-reads, from the cache, about 100,000 tokens of context. Projet Celsius estimate.
- "Re-read from cache" accounts for 55% of emissions in this scenario.
- 100% of the calculation rests on database assumptions; it does not measure actual use.
- Out of scope: server manufacturing and model training.
Sentence to citeRunning a coding agent, 1 hour, Projet Celsius estimate: 310 gCO2e per hour, for 670 Wh of electricity, servers worldwide (458 gCO2e/kWh).
Breakdown of emissions by item
310 gCO2e in total: 670 Wh of electricity × 458 gCO2e/kWh (servers worldwide).
Inputs of the calculation
PUE of the data centre, Re-read from cache, Read 1,000 tokens, Write 1,000 tokens
How the calculation is built
Built on the logs of intensive use published by the climatologist Zeke Hausfather (0.6 kWh for the median session, at his own factors); that a session lasts one hour is a database assumption, no source publishes the duration of a session. The tokens are not scaled to the calibration of the quick question (convention C9), which does not describe the work of an agent; the range comes from the cost of context re-read from the cache, which nobody publishes: 1% to 25% of a token read, around the 10% selected (convention C4).
Comparison with the other uses
↑ The most frugal use in the topic, "Code suggestion": 0.031 gCO2e, 9,800 times less
Emissions of one use, linear scale, 2 of the 6 uses in the topic. For calculated values, servers worldwide.
Characteristics
| Unit | 1 hour |
|---|---|
| Default location | world, 458 gCO2e/kWh (Ember 2025, life cycle); 11 other locations to choose from |
| Counted | the agent's steps, context re-read from the cache included, and cooling of the data centre |
| Not counted | server manufacturing and model training |
| Status | estimated, 3 other published values |
All the characteristics
| Identifier | IA-0048, projetcelsius.com/en/ai/database/IA-0048/ |
|---|---|
| Version | 0.4 of 12 November 2026 |
| Category | Uses, a request to AI |
| Choice of the selected value | database calculation from the entries this one is made of, at the quantities of the unit |
| Sources | calculation of the database |
| Licence | CC BY 4.0 |
Usage parameters
Advanced settings
Result for the chosen case
For a carbon footprint
| Item | purchases of services (GHG Protocol scope 3, category 1); regulatory BEGES, category 4 |
|---|---|
| Quantity | 1 hour |
| Factor | 310 gCO2e per hour, IA-0048 version 0.4, based on the local grid (location-based), servers worldwide |
| Total | 310 gCO2e · not counted: server manufacturing and model training |
Breakdown by item
Quantity, unit value, contribution and nature of the data for each item
| Item | Quantity | Unit value | Contribution | Share | Nature of the data |
|---|---|---|---|---|---|
| Re-read from cache | × 11,500 | 0.032 Wh | 370 Wh | 55 % | assumption |
| Read 1,000 tokens | × 432 | 0.32 Wh | 140 Wh | 21 % | assumption |
| Write 1,000 tokens | × 108 | 0.96 Wh | 100 Wh | 15 % | assumption |
| PUE of the data centre | × 1.10 on the rest | 61 Wh | 9 % | convention |
Sources of the calculation
3 publications, for this entry and those used to calculate it, with the chosen parameters.
The result is calculated by Projet Celsius. The sources below support its uses, its factors and its conversion; they do not all measure this total.
Energy and factors
The references of the selected values and of the entries used in this calculation.
Hausfather · Aug. 2026
Role and limits : Hausfather
Author. Zeke Hausfather, The Climate Brink
What the publication covers. intensive personal use of Claude Code, tokens recorded and converted into electricity
Reading by the database. A researcher's personal measurements, not a study
Coding agent hour → : Re-read from cache →
The context re-read from cache at each of the 120 steps of the hour: 96,000 tokens, 96% of a call in the median published session
Scope of the figure used. server, convention C4
Limit. This figure covers part of the calculation; it does not measure the whole entry.
Coding agent hour → : Read 1,000 tokens →
The new tokens of each step (files read, command results): 3,600, 3.6% of a call
Scope of the figure used. server, excluding the data centre
Limit. This figure covers part of the calculation; it does not measure the whole entry.
Coding agent hour → : Write 1,000 tokens →
The tokens written at each step, draft included: 900 on average in the published logs
Scope of the figure used. server, excluding the data centre
Limit. This figure covers part of the calculation; it does not measure the whole entry.
Watershed · Aug. 2026
Estimating GHG Emissions from AI Use: Framework for Corporate-Level Measurement ↗ (new tab)
Role and limits : Watershed
Author. Bistline, Ulissi, Davis, Glidden, Joyce, Li, Mohsenin and Suh (Watershed)
What the publication covers. factors per 1,000 tokens read and written at the server before PUE (host and standby machines included), PUE separate (1.10 for a large operator, 1.56 for an unknown operator)
What it does not cover. clean measurements, the model and the hardware
Reading by the database. Default factors proposed for review (public data as of June 2026); for a large operator, taken from Google's measurement published in 2025 (Elsworth et al.), which already contains the data centre overhead; constant per token, they underestimate long contexts and agents according to the authors
Reference for the variant "Medium model, Watershed default factor"
Scope of the figure used. Default factor for 1,000 tokens read, medium model, large operator, at the server before PUE (Watershed table 2, public data as of June 2026)
Limit. A default value that Watershed derives from Google's measurement (0.24 Wh for the median text query of the Gemini app), with about 500 tokens per query and a written token costing 3 times a read token; Watershed does not give the split between read and written tokens. The 0.24 Wh already include the overhead of Google's data centre (0.02 Wh, 8%): with 375 tokens read and 125 written, the simplest reading, the factor contains it, and the PUE of 1.10 applied afterwards counts it a second time (about 9%). Values proposed for review, final at the next version of the framework
Reference for the variant "Medium model, Watershed default factor"
Scope of the figure used. Default factor for 1,000 tokens written, medium model, large operator, at the server before PUE (Watershed table 2, public data as of June 2026)
Limit. 3 times the reading factor, based on the ratio between writing and reading selected by Watershed (Patel et al., Splitwise, 2024). Same derivation and same caveat as for reading: Google's 0.24 Wh already include about 8% of data centre. Values proposed for review, final at the next version of the framework
Coding agent hour → : Data centre PUE →
The data centre, convention C3
Scope of the figure used. building cooling and losses
Limit. This figure covers part of the calculation; it does not measure the whole entry.
Reference of the selected value
Scope of the figure used. Default factor for a large operator, from the server to the data centre; 1.56 for an unknown operator
Limit. A default value, not a measurement
Conversion into carbon
The source of the electricity factor of the selected location.
Worldwide: 458 gCO2e/kWh, life cycle.
Role and limits : Ember
Author. Ember, Our World in Data
What the publication covers. electricity generation over the life cycle, by country
What it does not cover. guarantees of origin
Reading by the database. The same source for all countries
Conversion of electricity into carbon: worldwide
Scope of the figure used. the production of electricity over the life cycle, including power plant construction and fuels
Limit. An average for the chosen country, not the electricity contract or the data centre of a specific service.
Published values
3 figures for the same operation, each within its own scope.
| Source | Date | Electricity | What the figure measures | Scope | Status |
|---|---|---|---|---|---|
| Coding agent hourselected by the database, v0.4 | 2026 | 672 Wh | the value selected by the database | same scope | estimated |
| Zeke Hausfather, The Climate Brink ↗The real energy use of agentic AI | August 2026 | 600 Wh | Median Claude Code session over 8 weeks of intensive use: about 10 million tokens and around a hundred calls to the model | other scope | estimated |
| Zeke Hausfather, The Climate Brink ↗The real energy use of agentic AI | August 2026 | 150 Wh | Claude Code, one user for 8 weeks, 1,138 instructions, context re-read from the cache at 96%, per instruction typed | other scope | estimated |
| Database checks and calculations | |||||
| Zeke Hausfather, Projet Celsius calculation ↗The real energy use of agentic AI | August 2026 | 12.1 Wh | About 170 kWh for more than 14,000 calls to the model, per call | other scope | computation |
The published figures do not all count the same scope; the scope of each figure is written next to it. Why do the estimates diverge?
Nature of the data
Published readings, conventions, assumptions
- Projet Celsius assumptions 100 %
Calculation assumptions
Quantities, nature and source of each assumption
Re-read from cache : × 11,500
The context re-read from cache at each of the 120 steps of the hour: 96,000 tokens, 96% of a call in the median published session
Read 1,000 tokens : × 432
The new tokens of each step (files read, command results): 3,600, 3.6% of a call
Write 1,000 tokens : × 108
The tokens written at each step, draft included: 900 on average in the published logs
Methodology conventions
With their alternatives
| The data centre | Server electricity is multiplied by 1.10 for the data centre |
|---|---|
| The cache | A token re-read from the cache costs 10% of a token read |
| The default location | Without other information, the servers are in the world, at the world average electricity |
| Electricity, location-based, on a life cycle basis | The kilowatt-hour is that of the servers' country, on a life cycle basis, from a single source for all countries |
| Two significant figures | A calculated value is rounded to two significant figures, a published value is reproduced with the precision of its source; the calculation keeps full precision |
| Calibration against production measurements | The tokens of the tasks are multiplied by a coefficient: the quick question recomposed from the machine's factors is brought back to its selected value, that of the consensus of published figures (one production measurement by Google, two estimates by Epoch AI and Microsoft, one statement by OpenAI). The code agent hour is not calibrated: its tokens, mostly context re-read from the cache, do not resemble those of a question |
| The formula | the sum of what makes up the entry, then CO2e: electricity (Wh) multiplied by the carbon intensity of the server location (gCO2e/kWh), divided by 1,000 |
| Calculate a case without an entry | The formula, the default values and a step-by-step example |
History
Entry in version 0.4 of 12 November 2026, in the database in version 0.4
- 12 November 2026Selected value : 440 Wh becomes 670 Wh (from 310 to 1,280 Wh).Decision of 12 November 2026: the C9 calibration comes from the token split of a question (100 read, 400 written) and does not describe an agent, of whose electricity 60% is context re-read from the cache, already counted by a convention (C4); Watershed and Microsoft write that constant per-token factors underestimate agents. The recipe is no longer calibrated. Range at the bounds of C4 (1% and 25% of a token read).
- 12 November 2026Description : scaled like all tasks to the database calibration: 0.4 kWh for the median session becomes the duration of one hour for a session is stated as a database assumption; the tokens are not calibrated (C9); the range comes from the cost of the cache (C4).No source publishes the duration of a session.
- 12 November 2026Version : 0.3 becomes 0.4.The selected value changes.
- 12 November 2026Zeke Hausfather's measurement, gap : the database recipe gives 0.62 kWh for 110 steps before calibration, 0.4 kWh after becomes the database recipe, without calibration, gives about 0.67 kWh for an hour of 120 steps.Same decision as the selected value of the agent hour.
- 2 November 2026Designation : A coding agent hour becomes Coding agent hour.More precise designation: the object first, the specification after the comma, the quantity or period at the end; the former title remains searchable.
- 2 November 2026Designation : Coding agent hour becomes Running a coding agent, 1 hour.The name in people's own words, without a leading article.
- 2 November 2026Short title : Agent hour becomes Coding agent hour.In people's own words.
- 2 November 2026Calculation : awaiting its value becomes database recipe.Composed in tokens.
- 2 November 2026Status : unpublished becomes estimated.Block composed with assumptions; it is calculated and published.
- 31 October 2026Creation of the database, in its first version (0.1).
