Claude-Carbon Estimates LLM Emissions, Details Methodology

This review examines claude-carbon, an open-source tool and its methodology for estimating CO2 emissions from Claude Code sessions. It highlights the significant energy cost of output tokens.

The Answer Up Front

For developers and organizations using Claude and seeking to understand or reduce their AI carbon footprint, claude-carbon offers a valuable, transparent methodology for estimating CO2 emissions. It is particularly useful for those who prioritize sustainability metrics alongside operational costs. Users who require official vendor-provided data or need to benchmark other LLMs beyond Claude's ecosystem should look elsewhere, as this tool's core data is specific to Claude 3.7 Sonnet. The bottom line is that claude-carbon provides a robust, if extrapolated, framework that clearly demonstrates the disproportionate energy cost of generating output tokens.

Methodology

This v0 review draws on the founder gwittebolle's published claims at dev.to, accessed on July 20, 2026. Independent benchmarks are pending. Update cadence: re-tested when claims diverge from observed behavior.

The tool under review is claude-carbon, an open-source project detailed in the provided source signal. This review covers the founder's methodology for estimating CO2e emissions from Claude Code sessions, including the foundational academic paper, the derived energy-per-token coefficients, and the cross-verification against EcoLogits. It also addresses the extrapolation strategy for other Claude models.

What is not covered in this review includes independent performance verification of the claude-carbon tool itself, long-term workflow integration, or edge-case emission scenarios. Our analysis relies entirely on the data and claims presented in the source article and the linked academic paper.

What It Does

Quantifying LLM Carbon Footprint

claude-carbon is an open-source tool designed to estimate the CO2e emissions generated by using Claude Code sessions. The founder, gwittebolle, developed it to provide a tangible number for the environmental impact of their daily LLM usage, much like a session cost. The tool reports personal usage data, such as 893 sessions resulting in roughly 970 kg CO2e over six months, with a median session around 200 g and a peak of 68 kg.

The Jegham et al. 2025 Foundation

The entire estimation hinges on a single academic paper: Jegham et al. 2025, "How Hungry is AI?" (arxiv.org/abs/2505.09598). This paper measures per-query inference energy for various models on AWS, specifically providing three configurations for Claude 3.7 Sonnet (PUE included):

From these three data points, gwittebolle performs an ordinary least-squares regression through the origin, yielding approximately 1.35e-4 Wh per input token and 2.88e-3 Wh per output token. This fit reportedly matches the higher-token configurations within 1%, which is relevant for typical coding sessions.

Carbon Intensity Conversion and Key Ratios

The energy values are converted to CO2e using the AWS region carbon intensity of 0.287 kgCO2e/kWh, a location-based figure cited by Jegham et al. This conversion results in the core emission figures:

• 39 gCO2e per million input tokens
• 826 gCO2e per million output tokens

This establishes a 21:1 ratio of output token emissions to input token emissions. The founder explains this disparity by noting that while FLOPs per token are similar in prefill and decode, hardware utilization differs. Prefill (input processing) saturates compute, whereas decode (output generation) is memory-bandwidth-bound, leading to higher energy consumption per token. The exact multiple depends on factors like batch size and serving stack.

Cross-Verification and Extrapolation

To corroborate these figures, the founder cross-checked against EcoLogits (ecologits.ai), which uses an independent methodology (parameter counts and a lifecycle model). EcoLogits' estimated range for Sonnet is ~565-1385 gCO2e per million output tokens, with claude-carbon's 826 gCO2e falling within this band. This independent corroboration strengthens the estimate.

For other Claude models like Opus, the tool extrapolates based on weaker signals, such as EcoLogits' parameter assumptions and Anthropic's price ratios, which both suggest a 1.7x-2x multiplier over Sonnet. The tool currently uses a 2x multiplier for Opus, acknowledging the inherent uncertainty in these extrapolated figures.

What's Interesting / What's Not

What makes claude-carbon particularly interesting is its explicit, open-source methodology. The founder's commitment to transparency, detailing the derivation from a specific academic paper and acknowledging the limitations, sets a high bar for environmental impact estimation in the LLM space. The core insight regarding the 21:1 ratio of output to input token CO2e, coupled with the explanation of hardware utilization differences (compute-bound prefill versus memory-bandwidth-bound decode), is a meaningful contribution. This highlights a critical, often overlooked, aspect of LLM efficiency that goes beyond simple token counts or API costs. The cross-verification with EcoLogits, an independent service, also adds significant credibility to the derived numbers, moving them beyond mere conjecture.

What is less compelling is the inherent fragility of extrapolating these measured Sonnet figures to other Claude models like Opus and Haiku. While necessary given the lack of official data, relying on parameter counts and pricing ratios for such critical environmental metrics introduces considerable uncertainty. The absence of official per-inference energy data from Anthropic or other frontier labs remains a significant gap in the industry, forcing developers like gwittebolle to construct these estimates. While claude-carbon provides a valuable framework, its specificity to Claude and reliance on a single academic source for foundational measurements mean it is not a universal LLM carbon calculator, nor does it offer independently verified performance claims for its own calculations.

Pricing

The claude-carbon tool is open-source and free to use. The blog post discusses the CO2e cost of using Claude, not a monetary price for the tool itself. Pricing snapshot date: July 20, 2026.

Verdict

claude-carbon is a highly recommended tool for Claude users who are serious about understanding and quantifying their environmental impact. Its strength lies in its transparent, data-driven methodology, which, despite relying on a single academic source for its core measurements, provides a credible estimate. The tool's most significant contribution is illuminating the disproportionate CO2e cost of generating output tokens compared to processing input tokens. This insight is crucial for optimizing LLM usage for both cost and sustainability. It is not suitable for those requiring official, vendor-supplied carbon data or for benchmarking non-Claude LLMs without further, more speculative, extrapolation.

What We'd Test Next

Our next steps would involve independently replicating the core energy measurements from the Jegham et al. 2025 paper for Claude 3.7 Sonnet, if feasible, to verify the foundational data. We would also benchmark claude-carbon's estimates against actual Claude API usage across a diverse range of prompt and response lengths to assess its accuracy in real-world scenarios. Further investigation into the energy profiles of other major LLMs (e.g., GPT, Gemini, Llama) using a similar, transparent methodology would be valuable. We would also explore how different batch sizes and serving stack configurations impact the observed input/output token energy ratio, and rigorously evaluate the accuracy of the extrapolated numbers for Claude Opus and Haiku through dedicated empirical testing.

The investor read

The emergence of tools like claude-carbon signals a growing market demand for transparency and accountability in AI's environmental impact, moving beyond purely performance and cost metrics. As ESG (Environmental, Social, and Governance) reporting becomes more stringent and regulatory pressures increase, companies will need verifiable methods to quantify their AI carbon footprint. The significant 21:1 input:output token CO2e ratio highlighted by this tool suggests that optimizing LLM output efficiency will be a critical area for innovation and investment, impacting both operational costs and sustainability goals. While claude-carbon itself is an open-source, bootstrapped effort, it points to a future where dedicated AI carbon accounting platforms, offering verified data and enterprise integrations, could become highly investable, especially if they can generalize across multiple LLM providers and integrate with existing sustainability reporting frameworks. EcoLogits is a comparable player in this nascent space.