Faire une recherche approfondie

Name: Carry out deep research
Creator: Projet Celsius
License: https://creativecommons.org/licenses/by/4.0/

Projet Celsius

Database entryUsesLevel 3 of 6 useUse, level 3 of 6Used to calculate 17 entries, level 4Calculated with 6 entries, levels 1 to 3Database structure

Carry out deep research

The assistant launches dozens of searches, reads about a hundred web pages over several minutes, then writes a report of 5 pages. Projet Celsius estimate.

"Read 1,000 tokens" accounts for 34% of emissions in this scenario.
100% of the calculation rests on database assumptions; it does not measure actual use.
Out of scope: server manufacturing, model training, the network and the user's device.

Compare

Calculated with

Used to calculate

Breakdown of emissions by item

45 gCO2e in total: 99 Wh of electricity × 458 gCO2e/kWh (servers worldwide).

Inputs of the calculation 5

PUE of the data centre, Read 1,000 tokens, Re-read from cache, Write 1,000 tokens, Web search (benchmark)

Annual emissions for carrying out deep research, by frequency

Your frequencyPeriod

Read the values at10 times a year

Reading: a single use of "Deep research" weighs as much as 5.2 phone charges. Carbon footprint (CO2e) over one year, by number of uses. Each use is a straight line; "Deep research" is highlighted, with the chosen parameters, compared with 5 reference uses at their default settings; the solid line shows one phone charge per week over a year, and the point where the task reaches it. Servers worldwide. Calculation from the AI Impact Database by Projet Celsius.

How the calculation is built

Composed in tokens; compar:IA gives 94 Wh for writing the report alone, with larger models: on this same scope, the database counts about 16 Wh.

Comparison with the other uses

The topic "Research"

Literature search13 gCO2e, 0.3 times the value of "Deep research"
Deep research45 gCO2e

↑ The most frugal use in the topic, "AI overview of a search": 0.061 gCO2e, 750 times less

Emissions of one use, linear scale, 2 of the 5 uses in the topic, those of the same order of magnitude (at most 25 times between the smallest and the largest value). For calculated values, servers worldwide.

Characteristics

Unit	1 search
Default location	world, 458 gCO2e/kWh (Ember 2025, life cycle); 11 other locations to choose from
Counted	the searches, the reading of pages, the context re-read at each step, the writing of the report and cooling of the data centre
Not counted	server manufacturing, model training, the network and the user's device
Status	estimated, 1 other published value

All the characteristics identifier, version, rule for the value, sources, licence

Identifier	IA-0045, projetcelsius.com/en/ai/database/IA-0045/
Version	0.3 of 2 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

Emissions by server location

Emissions associated with electricity (CO2e)

Poland58 gCO2e
World45 gCO2e
United States38 gCO2e
European Union21 gCO2e
France4.1 gCO2e

× 14 between Poland and France, for the same electricity

Ember 2025, over the life cycle. The same use consumes 99 Wh in each location.

Carrying out deep research · result per search · estimated · one server location per line.Counted: the searches, the reading of pages, the context re-read at each step, the writing of the report and cooling of the data centre. Out of scope: server manufacturing, model training, the network and the user's deviceFind this case: projetcelsius.com/en/ai/database/tasks/deep-research/Projet Celsius calculation, AI Impact Database, version 0.4 of 13 November 2026 · IA-0045, entry added in version 0.3 · engine 0.3.1 · CC BY 4.0.

Usage parameters

Advanced settings model class, calibration

Model classCalibration

Result for the chosen case

Location of the servers

For a carbon footprint

The row for a carbon footprint

Item	purchases of services (GHG Protocol scope 3, category 1); regulatory BEGES, category 4
Quantity	1 search
Factor	45 gCO2e per search, IA-0045 version 0.3, based on the local grid (location-based), servers worldwide
Total	45 gCO2e · not counted: server manufacturing, model training, the network and the user's device

Breakdown of emissions by item

Emissions associated with electricity (CO2e)

Read 1,000 tokens15 gCO2e · 34 %
Re-read from cache15 gCO2e · 34 %
Write 1,000 tokens6.7 gCO2e · 15 %
PUE of the data centre4.1 gCO2e · 9.1 %
Web search (benchmark)3.7 gCO2e · 8.3 %

Carrying out deep research · result per search · estimated · servers worldwide.Counted: the searches, the reading of pages, the context re-read at each step, the writing of the report and cooling of the data centre. Out of scope: server manufacturing, model training, the network and the user's deviceFind this case: projetcelsius.com/en/ai/database/tasks/deep-research/Projet Celsius calculation, AI Impact Database, version 0.4 of 13 November 2026 · IA-0045, entry added in version 0.3 · engine 0.3.1 · CC BY 4.0.

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
Read 1,000 tokens	× 105 after calibration, 160 before	0.32 Wh	34 Wh	34 %	assumption · Google
Re-read from cache	× 1,050 after calibration, 1,600 before	0.032 Wh	34 Wh	34 %	assumption · Anthropic
Write 1,000 tokens	× 15.2 after calibration, 23.2 before	0.96 Wh	15 Wh	15 %	assumption
Web search (benchmark)	× 27.3	0.3 Wh	8.2 Wh	8 %	assumption
PUE of the data centre	× 1.10 on the rest		9 Wh	9 %	convention

Entries that use it (17)

End-to-end study, Automated monitoring, 1 year, Manager or executive, 1 year, Consultant, 1 year, Researcher, 1 year, Salesperson, 1 year, Student, 1 year, Free personal use, 1 year and

Sources of the calculation

6 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.

Uses and assumptions

These publications shed light on AI tasks and uses. They do not measure the carbon footprint of the case calculated.

Anthropic · June 2025
How we built our multi-agent research system ↗ (new tab)
Re-read from cache
Role and limits : Anthropic
What the publication covers. an agent uses about 4 times the tokens of a conversation, a multi-agent system about 15 times
What it does not cover. electricity
Reading by the database. Publisher orders of magnitude, with no absolute value
- Deep research → : Re-read from cache →
  The context re-read from cache at each of the 40 steps, 40,000 tokens on average
  Scope of the figure used. server, convention C4
  Limit. The quantity and the share entrusted to AI are database assumptions, not a measured footprint.
Google · 2026
Gemini Deep Research ↗ (new tab)
Read 1,000 tokens
Role and limits : Google
What the publication covers. up to hundreds of sites browsed and pages processed per search
What it does not cover. electricity
Reading by the database. A presentation page: a ceiling, not an average
- Deep research → : Read 1,000 tokens →
  The web pages read, about a hundred by default: Google announces up to hundreds of sites
  Scope of the figure used. server, excluding the data centre
  Limit. The quantity and the share entrusted to AI are database assumptions, not a measured footprint.
Projet Celsius · November 2026
Tokens per word, per line of code and per cell, measured with 7 public tokenisers
Write 1,000 tokens
Role and limits : Projet Celsius
What the publication covers. 479,888 words from 196 blog articles and their English translation, code from the site and the Python library, a table of measurements; tokenisers of GPT-4o, GPT-4, Llama 3, Gemini (via Gemma 3), Mistral, DeepSeek and Qwen
What it does not cover. the Claude tokeniser, not public
Reading by the database. A measurement on technical texts, figures included; to be repeated if publishers change their tokeniser
Projet Celsius measurement, repeated on request with the public tokenisers
- Deep research → : Write 1,000 tokens →
  Length convention: one page
  Scope of the figure used. A report page, headings and paragraphs included: between the 1,500-character sheet (250 words) and a full A4 page in single spacing (about 550 words, at 6.1 characters per word measured on our articles)
  Limit. This publication supports the quantity of text or content processed, not its carbon footprint.

Energy and factors

The references of the selected values and of the entries used in this calculation.

Google · January 2009
Powering a Google search ↗ (new tab)
Web search (benchmark)
Role and limits : Google
What the publication covers. a search, data centre
Reading by the database. Never remeasured
- Web search (reference) →
  Reference of the selected value
  Scope of the figure used. A search, data centre included
  Limit. Never publicly remeasured; Google's AI answers are not published
Watershed · Aug. 2026
Estimating GHG Emissions from AI Use: Framework for Corporate-Level Measurement ↗ (new tab)
Read 1,000 tokens · Write 1,000 tokens · Data centre PUE
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
- Read 1,000 tokens →
  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
- Write 1,000 tokens →
  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
- Deep research → : 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.
- Data centre PUE →
  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.

Ember · 2026
Global Electricity Review 2026, and Our World in Data based on Ember ↗ (new tab)
Grid electricity
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
- Grid electricity →
  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

1 figure for the same operation, each within its own scope.

Green: same scope as the selected value (measured in dark, estimated or published in light). Grey: other scope. Empty circle: check by the database or discarded figure. Logarithmic scale.

Source	Date	Electricity	What the figure measures	Scope	Status
Deep researchselected by the database, v0.3	2026	98.9 Wh	the value selected by the database	same scope	estimated
compar:IA ↗compar:IA	2026	94 Wh	Median, model output only · The pages read are not counted	check	estimated

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

What this figure rests on

Projet Celsius assumptions 100 %

A share counts as an assumption as soon as a quantity set by Projet Celsius (a frequency, a share done with AI, a headcount) enters its calculation, even if the rest rests on published measurements.

Calculation assumptions

Quantities, nature and source of each assumption

Web search (benchmark) : × 27.3
The search engine queries, 30 by default, counted as conventional searches; divided by 1.10, as the conventional search already counts its data centre
assumption · Web search (benchmark)
Read 1,000 tokens : × 105
The web pages read, about a hundred by default: Google announces up to hundreds of sites
assumption · Google · Read 1,000 tokens
Re-read from cache : × 1,050
The context re-read from cache at each of the 40 steps, 40,000 tokens on average
assumption · Anthropic · Re-read from cache
Write 1,000 tokens : × 15.2
The draft of each step, 500 tokens, then the report of 5 pages
assumption · Write 1,000 tokens

Methodology conventions

With their alternatives

The full methodology

The data centre	Server electricity is multiplied by 1.10 for the data centre · alternatives: × 1.72 from chips to the full service at Google, never to be combined
The cache	A token re-read from the cache costs 10% of a token read · alternatives: bounds 1% and 25%
The default location	Without other information, the servers are in the world, at the world average electricity · alternatives: the United States, where most of the large assistants run
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 · alternatives: market-based, with purchase contracts, shown alongside when published
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
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 0.3 of 2 November 2026, in the database in version 0.4

The changelog

2 November 2026Designation : A deep research becomes Deep research.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 : Deep research becomes Carrying out deep research.In people's own words.
2 November 2026Address : recherche-approfondie becomes faire-une-recherche-approfondie.In people's own words; the old address redirected.
2 November 2026Calculation : recorded value becomes database recipe.Composed in tokens.
2 November 2026Selected value (94).The value is now calculated by the database, from tokens.
2 November 2026Status : working value becomes estimated.Block composed with assumptions.
31 October 2026Creation of the database, in its first version (0.1).

Citation

Projet Celsius, AI Impact Database, version 0.4 of 13 November 2026, "Carrying out deep research", IA-0045 (entry added in version 0.3 of 2 November 2026), projetcelsius.com/en/ai/database/IA-0045/ · CC BY 4.0.

Download

JSON Markdown