What is Claude's carbon footprint?
Which model and which use weigh the most?

A short conversation with Claude Haiku 4.5 uses about 0.08 Wh (watt-hour) and emits 0.03 gCO2e, the same as a 10 W LED bulb left on for 30 seconds. With Claude Opus 5, the same conversation uses 20 times more, 1.6 Wh. And with that same Opus 5, an agent session that codes for 1 to 2 hours uses 43 Wh, as much as 27 conversations.

These are estimates: Anthropic publishes neither the energy use nor the size of its models, and the only database that estimates them version by version, CLEER (Closed-model Latent Energy Estimation Range), was published on 29 September 2026 by independent researchers. Intelligence does not track energy: GPT-5.6 Luna, which scores like Sonnet 5, is estimated to be 5 times more frugal, although that estimate is extrapolated. The sections below go through these figures model by model, then look at what makes them vary.

1How much energy does each Claude model use?

The CLEER database is published by the Sustainable AI Group, around Nidhal Jegham, Chan Young Koh, Boris Gamazaychikov and Sasha Luccioni, one of the most cited researchers on the energy use of AI. For a closed model such as Claude, it estimates the size of the model from its knowledge score, its context window and its release date. It then matches it to 5 open models measured on recent graphics cards, and calibrates it against its public output speed.

The result covers the total energy of an average US data centre, cooling included, for the same session and the same number of tokens whatever the model.

From Haiku to Fable, a factor of 35

CLEER gives a value for each version of Claude, based on a typical conversation: one question, one answer, then 1 or 2 follow-ups. The latest estimated version of each model family ranks as follows:

• Claude Haiku 4.5, released in October 2025, uses 0.08 Wh; CLEER matches it to open models with 20 to 120 billion parameters in total, of which 3 to 5 billion are active.
• Claude Sonnet 5 uses 0.83 Wh, after 0.74 Wh for Sonnet 4, 1.20 Wh for Sonnet 4.5 and 0.72 Wh for Sonnet 4.6.
• Claude Opus 5 uses 1.61 Wh, within a range of 1.37 to 1.88 Wh across the 7 versions of Opus released since May 2025.
• Claude Fable 5 uses 2.80 Wh, an extrapolated value, because Fable outstrips every open model used as a reference.

Opus 5 therefore uses about 2 times as much as Sonnet 5 and 20 times as much as Haiku 4.5. The current models, Haiku 5.5, Sonnet 5.5, Opus 5.5 and Fable 5.1, released in September and October 2026, have no estimate yet: CLEER updates its data every quarter. For the orders of magnitude of AI in general, from the single query to the data centre, our article on the carbon footprint of AI is the place to start.

What about CO2 and water?

The CO2 depends on the electricity that powers the data centre. Average US electricity emits 384 gCO2e per kWh over its life cycle in 2025, according to Our World in Data. A conversation then emits from 0.03 gCO2e with Haiku 4.5 to 1.1 gCO2e with Fable 5, close to 8 metres in a petrol or diesel car. On French electricity, at 41 g per kWh, it emits 9 times less. By default, CLEER uses a heavier case, gas turbines installed on site, at 640 g per kWh.

Water follows electricity. CLEER does not quantify it. Cooling in US data centres used an average of 0.36 litre per kWh of IT energy up to 2023, and 0.45 to 0.48 since, according to the Lawrence Berkeley National Laboratory. With the low value, our calculation gives from 0.02 mL with Haiku 4.5 to 0.7 mL with Fable 5 per conversation, the same order as the 0.26 mL Google measured for a question to Gemini.

The water used by the power plants that generate this electricity weighs more than 10 times as much, about 4.5 litres per kWh according to the same laboratory: our article on the water footprint of AI explains these 2 scopes.

How reliable these estimates are

CLEER gives each model a confidence score from 0 to 1: high above 0.85, good from 0.55 to 0.85, lower from 0.30 to 0.55, extrapolated below that. Haiku 4.5 (0.68) and Sonnet 5 (0.75) are well estimated, Opus 5 (0.47) less so, and Fable 5 (0.13) remains an extrapolation. The database has not been peer reviewed.

The study "How Hungry is AI?", co-authored in 2025 by Nidhal Jegham, gave 0.95 Wh for a short query to Claude 3.7 Sonnet, and slightly more for Claude 3.5 Haiku, for lack of a size assumption specific to each model. The figures that circulate for "Claude" therefore mix model families whose energy use spans a ratio of 1 to 35.

2Why does Opus use more energy than Haiku?

Producing 1 token takes about 2 computing operations per active parameter of the model, according to Epoch AI, and electricity follows computation. The energy of an answer therefore comes down to a product: the energy per token, set by the model and its hardware, multiplied by the number of tokens processed, set by how it is used. On the first term, Anthropic publishes neither the size of its models, nor their architecture, nor their energy use.

Model size, which nobody publishes

The model cards in Anthropic's Transparency Hub describe training and safety testing, without any figure on size, energy or water. The EcoLogits method assumed about 2 trillion parameters for Claude 3 Opus, of which 200 to 600 billion are active for each token, based on a "mixture of experts" architecture that activates only part of the network. At the other end of the range, CLEER matches Haiku 4.5 to open models of 20 to 120 billion parameters in total, of which only 3 to 5 billion are active.

Output speed, an imperfect clue

Generation speed gives an indirect clue. According to measurements by Artificial Analysis in early November 2026, Haiku 5.5 produces about 3 times as many tokens per second as Fable 5.1, and Sonnet 5.5 slightly more than Opus 5.5. A slower model generally does more computation per token, but speed also depends on how the model is served: Anthropic sells a more expensive fast mode for Opus, which serves the same model faster.

Hardware and request batches

Claude runs on 3 families of chips. Anthropic has announced the use of up to 1 million Google TPUs and more than 1 million AWS Trainium2 chips, and rents NVIDIA cards in Microsoft data centres. The split between them is not published.

The number of requests processed together, called a batch, also matters: a larger batch spreads the fixed cost of the machines and narrows the efficiency gap between small and large models. Our article on why AI is bad for the environment traces these chips back to their manufacture.

These efficiencies are improving fast, but a recent model is not necessarily more frugal than its predecessor. Of 14 recent models tested by Hugging Face's AI Energy Score, 8 used at least as much energy as a model of similar size tested in February 2025.

3Why does an AI agent use so much energy?

An agent such as Claude Code does not answer once. It reads files, runs commands, checks its work, and each step sends the model the whole conversation from the start. The Claude Code documentation says so: a one-line question, in a session left open all day, uses tokens for the whole conversation. As early as June 2025, Anthropic measured that an agent spends about 4 times the tokens of a conversation, and a multi-agent system about 15 times.

The diagram reads in 2 directions. For the same use, moving from Haiku 4.5 to Fable 5 multiplies energy by 35. With the same model, moving from a single question to a day of agents multiplies it by nearly 370: from 0.46 Wh to 167 Wh for Opus 5, a 10 W LED bulb left on for 17 hours.

According to CLEER, a typical agent session uses 27 to 67 times as much as a conversation depending on the model, and a day of agents more than 100 times: in proportion, use weighs most on the small models.

Context reread at every step

Reread tokens make up most of the count. Over 8 weeks of heavy Claude Code use, the climate scientist Zeke Hausfather counted 3.2 billion tokens. 96% were cache reads, context already seen that the model rereads at each call, and 0.4% was generated text.

Anthropic charges these reads at 10% of the input price, 5% on Opus 5.5 and Sonnet 5.5, and 2.5% on Fable 5.1: cheaper per token, they make up most of the volume. Their weight in energy rests on an assumption, that a reread token costs about 10% of the energy of a new token, with bounds of 1 to 25% in Hausfather's work.

Length also matters for a single query. In "How Hungry is AI?", moving from a short query, 100 input tokens and 300 output tokens, to a long one, 10,000 and 1,500, multiplies by 6 the estimated energy of a query to Claude 3.7 Sonnet. The context window of current models reaches 1 million tokens, and Anthropic multiplies the price of Haiku 5.5 by 5 beyond 100,000 tokens. This role of tokens, rather than of the "query" alone, is explained in our article on AI electricity consumption.

What Claude Code users measure

The published measurements cover individual users, each with their own assumptions:

• Simon Couch, a developer, estimates his median Claude Code session at 41 Wh, 24 queries and 592,000 tokens, and his median day at 1.3 kWh.
• Zeke Hausfather arrives at about 150 Wh per typed instruction, 3 kWh for an average day and 370 kgCO2e per year, for heavy use on the US grid.
• Anthropic puts the average cost of Claude Code in companies at $13 per developer per active day, and below $30 for 90% of users.

Anthropic itself lists the habits that reduce this consumption. Its documentation recommends Sonnet for most coding tasks, Haiku for simple subagents, the "/clear" command between 2 unrelated tasks, and precise requests, since a vague request triggers an exploration of the whole codebase. It notes that unexpected spending generally comes from sessions that are never cleared or from Opus left as the default model. For a whole workstation, these figures can be read alongside our article on a company's digital carbon footprint.

4What difference does the effort level make?

Current Claude models think before they answer. On Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1, this thinking cannot be switched off, and its tokens are billed as output text, up to tens of thousands per query. According to Anthropic, the effort level, from "low" to "max", sets how many tokens Claude spends, thinking and tool calls included.

On the Artificial Analysis benchmark, the same task with Opus 5.5 costs about 10 times more at "max" effort than at "low" effort, from $0.98 to $10.04. The gap is the same for Haiku 5.5, reaches 20 times for Sonnet 5.5 and stays close to 3 times for Fable 5.1.

Since the price of a token is fixed for a given model, we read this gap as a gap in tokens, not in measured energy. Opus 5.5's intelligence index, for its part, rises from 35 to 58, on the current Artificial Analysis index, which is more recent than the one in the opening table.

The setting can even reverse the ranking of model families. Opus 5.5 at "low" effort cost less than Sonnet 5.5 at "high" effort, $0.98 against $1.64 per task, for a slightly lower index. By default, Anthropic sets Opus 5.5 and Haiku 5.5 to "medium", and Sonnet 5.5 and Fable 5.1 to "high": that is the setting a team keeps if it never changes it.

Thinking, tokens you do not see

Direct measurements point the same way. In the first version of "How Hungry is AI?", in May 2025, Claude 3.7 Sonnet with extended thinking used 2 to 4 times the energy of the same query without thinking.

Across 22 open models, the measurements published in May 2026 by Arcep, the French telecoms regulator, show 92% more energy use on average with reasoning mode on. The AI Energy Score, for its part, finds 154 to 697 times more for the same model, across 3 models and counting the graphics card only. Our article on the habits that reduce the impact of AI draws the practical rule: do not leave the most advanced reasoning on by default.

Each model's own verbosity

With the token count held equal, CLEER compares models, but each has its own way of answering. On the Artificial Analysis benchmark, Sonnet 5 produces nearly 3 times as many tokens as the median model for the same task, Opus 5 1.8 times and Haiku 4.5 less than half. With this verbosity, published by CLEER on 30 October 2026, the gap between Sonnet 5 and Opus 5 falls from 1.9 to 1.2 times, 2.3 Wh against 2.8 Wh per conversation. Choosing a more frugal AI is therefore a matter of testing on real tasks.

5Does a model's price reflect its energy use?

On the same models, price understates the energy gap: from Haiku 4.5 to Fable 5, the price of output tokens is multiplied by 10, from $5 to $50 per million, and estimated energy by 35. According to CLEER, in about 1 in 5 comparisons between 2 models, the cheaper one uses more energy.

Price reflects computation well for open models, sold by several competing hosting providers. For a closed model, CLEER's technical report states that this link no longer holds: the margin may reflect a subsidy, a commercial strategy or platform lock-in.

Opus, a price divided by 3

Opus is a case in point. From Opus 4.1 to Opus 4.5, in November 2025, Anthropic divided its price by 3, from $75 to $25 per million output tokens, while its estimated energy stayed around 1.5 Wh per conversation. Between Haiku 4.5 and Opus 4.5, price is multiplied by 5 and energy by 19. Among the closed models in CLEER, the energy bought with $1 ranges from 3 Wh for Haiku 4.5 to more than 40 Wh for several OpenAI models, according to our calculation.

Version after version, Opus uses almost the same energy, 1.37 to 1.88 Wh, while its intelligence index rises from 21 to 51. Part of this stability comes from the method, which infers size from the model's capabilities: capability progresses faster than estimated energy, without any measurement to confirm it. Since Opus 4.7, the new tokenizer counts about 35% more tokens for the same text: at an equal price per token, that text therefore costs more.

Between model families, a useful guide

Within the same generation, the most expensive family does more computation: Opus 5 costs 2.5 times the price of Sonnet 5 and uses about 2 times as much energy. Price remains a useful guide between families, much less so from one version to the next. Methods that convert dollars into kilowatt-hours, to quantify Claude Code from the invoice, inherit this limit. To count AI in a Bilan Carbone® (the French carbon accounting method), under scope 3 purchased services, token logs are better than invoices.

6What about ChatGPT, Gemini or Mistral?

Claude's competitors follow the same structure in 3 model families. CLEER finds a gap of nearly 4 times on average between the top family and the light one, less than the gap between the latest Claude versions (20 times from Haiku 4.5 to Opus 5).

For a typical conversation, GPT-5.6 Sol uses 1.36 Wh, GPT-5.6 Terra 0.55 Wh and GPT-5.6 Luna 0.17 Wh, and Gemini 3.7 Flash, calibrated on Google's chips, 0.50 Wh. Gemini's top families and Mistral's models are not in the database. Our comparison of ChatGPT, Claude and Gemini says which to choose for which use.

Mistral AI, July 2025

The life cycle assessment of an answer from Mistral Large 2

Mistral AI, Our contribution to a global environmental standard for AI, 22 July 2025. Document cited.

The only life cycle assessment published by an AI provider, carried out with ADEME (the French Agency for Ecological Transition) and reviewed by third parties. For a 400-token answer: 1.14 gCO2e and 45 mL of water, server manufacturing included. This scope, wider than electricity, cannot be compared with the rows of the opening table.

These 3 providers have, for their part, published a figure, each within its own scope:

• Google measured in production 0.24 Wh, 0.03 gCO2e and 0.26 mL of water for the median text prompt to Gemini, in May 2025, in a study by its engineers. It is the only provider measurement published.
• OpenAI stated 0.34 Wh and 0.32 mL for an average ChatGPT query, in June 2025, without a published method. Our article on the ChatGPT query traces its origin, and the one on ChatGPT versus a Google search compares the two.
• Mistral AI published the life cycle assessment of a 400-token answer from its Large 2 model: 1.14 gCO2e and 45 mL of water, server manufacturing included. This life cycle assessment (LCA) scope goes well beyond electricity.

The models in service in early November 2026, GPT-6 Astra and GPT-6.1 Sol at OpenAI, Gemini 4 Argon at Google, Opus 5.5 and Fable 5.1 at Anthropic, have no public estimate yet. These provider figures apply to a single question: for one question, CLEER estimates 0.14 Wh for Gemini 3.7 Flash, the same order as Google's measurement. For the same model family, ChatGPT and Gemini stay in the same order of magnitude as Claude, 0.1 to 3 Wh per conversation: as much as a generated image for the largest models, far less than a generated video.

At Projet Celsius, we use Claude among other tools. For us, the choice that matters is between uses, model families and effort levels, far more than between providers, which stay in the same order of magnitude for the same family. Leaving Opus at maximum effort, on agents that run all day, weighs more in a team's energy use than a change of provider. The useful step is to set a default model and effort level, Sonnet or Haiku for subagents, and to clear the context between 2 tasks.

The question to put to Anthropic is the one Google has already answered: the median energy of a query, model by model, measured in production.

7Key takeaways

• Anthropic publishes neither the energy use nor the size of its models: the only per-version values are independent estimates, those of CLEER in September 2026.
• A conversation emits from 0.03 gCO2e with Haiku 4.5 to 1.1 gCO2e with Fable 5 on the US grid, for 0.08 to 2.8 Wh; Sonnet 5 and Opus 5 are at 0.8 and 1.6 Wh.
• An agent rereads its whole context at every step: a typical session uses 27 to 67 times as much as a conversation, and a day of agents more than 100 times.
• Depending on the model, the effort level multiplies the tokens spent on the same task by 3 to 20, and Opus 5.5 at low effort spends less than Sonnet 5.5 at high effort.
• Price is a poor guide to energy: Opus's price was divided by 3 in November 2025 without its estimated consumption moving.

To place Claude within the full set of AI impacts, climate, electricity, water and metals, see our article on the environmental impact of AI, and to quantify your own use, the AI footprint tool.

Further resources

Frequently asked questions

1What is Claude's carbon footprint?

According to the CLEER estimates published in September 2026, a short conversation with Claude emits from 0.03 gCO2e with Haiku 4.5 to about 1.1 gCO2e with Fable 5, on average US electricity, and 9 times less on French electricity. A Claude Code agent session emits 2 to 30 gCO2e depending on the model. Anthropic publishes no figures on energy or emissions, and its September and October 2026 versions have not yet been estimated.

2Which Claude model uses the least energy?

Haiku, the light model family. For the same conversation, Claude Haiku 4.5 uses about 0.08 Wh according to CLEER, 10 times less than Sonnet 5 and 20 times less than Opus 5. The current versions, including Haiku 5.5, have no estimate yet. Anthropic itself recommends Haiku for simple tasks and subagents.

3How much energy does Claude Opus use?

A conversation with Claude Opus uses 1.4 to 1.9 Wh depending on the version, and 1.6 Wh for Opus 5, according to CLEER's September 2026 estimates. An agent session with Opus 5 uses about 43 Wh, and a working day of agents about 170 Wh. Opus 5.5, the current version, has not yet been estimated.

4Does Claude's extended thinking use more energy?

Yes, because it produces more tokens. On Claude 3.7 Sonnet, extended thinking multiplied the energy of a query by 2 to 4 according to the study "How Hungry is AI?". On current models, thinking is always on and the effort level sets how much of it there is: from "low" to "max", the same task spends about 10 times more tokens on Opus 5.5.

5What is the carbon footprint of Claude Code?

A Claude Code session uses tens of watt-hours: 41 Wh for a developer's median session according to Simon Couch, and 43 Wh for a typical agent session with Opus 5 according to CLEER, or about 17 gCO2e on the US grid. Heavy use reaches about 3 kWh per day and 370 kgCO2e per year, according to Zeke Hausfather. The model, the reread context and the effort level account for most of the gap.

6Does Claude use more energy than ChatGPT?

According to CLEER's estimates, they are in the same order of magnitude for the same model family: 1.6 Wh for Opus 5 and 1.4 Wh for GPT-5.6 Sol per conversation. At a similar intelligence index, the gap can be wider: GPT-5.6 Luna, which scores like Sonnet 5, is estimated at 0.17 Wh against 0.83 Wh, but by extrapolation. Use, agent or conversation, weighs more than the brand.

Ask Guillaume your questionA personal reply within 24 working hours

Send my questionor: guillaume@projetcelsius.com