Anthropic's Carbon Footprint: How Green Is Claude?

Anthropic is vocal about AI safety. On environmental impact, the company is much quieter. Unlike Google and Microsoft, Anthropic has not published an annual sustainability report with specific emissions data. That gap makes it hard to evaluate how green Claude actually is. This post pulls together what is publicly known, what can be inferred from infrastructure choices, and how per-query emissions compare to other frontier models.

Where Claude Runs

Claude runs almost entirely on AWS. Anthropic signed a $4 billion deal with Amazon in 2023, and the majority of API traffic is served from AWS data centers, primarily US East (Virginia) and US West (Oregon).

This matters because the carbon intensity of a query depends heavily on where the GPU is located:

• US East (Virginia): approximately 310 gCO2/kWh grid intensity, classified as high water stress by the WRI Aqueduct index
• US West (Oregon): approximately 118 gCO2/kWh, moderate water stress due to Columbia River hydro
• EU regions (Frankfurt, Ireland): 300 and 350 gCO2/kWh respectively

AWS has committed to 100% renewable energy matching by 2025 (they reached it in 2023), but renewable matching is not the same as running on clean power in real time. The Virginia grid still draws coal and gas during peak hours.

Estimated Emissions Per Claude Query

Anthropic has not disclosed per-token energy figures. The estimates below are based on published academic research, Google's published TPU energy data (which Claude uses via Google Cloud secondarily), and model architecture benchmarks from Hugging Face and MLCommons.

The takeaway: Claude Haiku is genuinely lightweight. Claude Opus is in the same range as GPT-4o. The gap between smallest and largest model in the Claude family is roughly 17x in energy per token.

How Anthropic Compares to OpenAI and Google

Google disclosed a 48% increase in total emissions between 2022 and 2024, explicitly attributed to AI infrastructure growth. Microsoft reported a 29% increase over the same period. Both companies still publish annual environmental reports with specific kWh and CO2 figures.

Anthropic has not done this. The company references commitments to responsible AI, but there is no public-facing page with fleet-level energy or emissions data as of mid-2026.

This is not unusual for a private company at Anthropic's stage. But it does make independent auditing harder. The best proxy is AWS's own sustainability reporting, which covers data center PUE and renewable matching at the infrastructure level, not the workload level.

What Actually Determines Claude's Footprint Per Request

Three variables dominate the per-query carbon number:

• Model size. Haiku versus Opus is a 17x difference. If your task does not need Opus reasoning depth, you are spending 17x the energy for nothing.
• Output token count. Output tokens cost roughly 3x more compute than input tokens. Long responses matter more than long prompts.
• Data center region. Routing the same Sonnet query through AWS Oregon instead of AWS Virginia cuts carbon by about 60% and removes it from a high water stress zone.

The Broader Picture

The UN UNU-INWEH report published in June 2026 projects AI data center water consumption equivalent to the domestic needs of 1.3 billion people by 2030 if current growth rates hold. Anthropic's infrastructure is a small share of that today, but the model is scaling fast. Claude API usage has grown sharply since the 2023 Amazon deal.

Until Anthropic publishes its own emissions data, the most useful thing developers can do is use the smaller model where possible, measure actual token volumes, and pick the right region for API calls. We built the AI Impact Calculator specifically for this.