Aurxo Baseline v1.0 · Evidence reviewed 27 July 2026

What the evidence supports.

Aurxo separates current product assessments from scoped model references. A dash is not a poor grade. It means the evidence is not complete or comparable enough for a universal product label.

Only Gemini receives a current product-level impact grade in this first baseline. The other rows make missing evidence and available reference studies visible instead of turning them into invented product scores.

Current product assessments

Product / scope

Grade

Confidence

What Aurxo can responsibly say

Gemini Apps

Median production text prompt · May 2025

Moderate

Google reports 0.24 Wh, 0.03 gCO₂e and 0.26 mL across the serving stack. This is a provider-authored production study, not an Aurxo audit.

Source

ChatGPT

Current routed product

—

Low

OpenAI reports average-query energy and water, but no comparable current product carbon value or reproducible routed-product boundary. A scoped historical GPT-4o reference appears below.

OpenAI evidence

Claude

Current routed product

—

Insufficient

Anthropic provides extensive model and safety disclosures, but Aurxo did not locate complete current product-level energy, carbon and water metrics. A Claude 3.7 reference appears below.

Source review

Microsoft Copilot

Current consumer product

—

Insufficient

Microsoft Research publishes valuable industry-level inference evidence, including a 0.31 Wh median for an optimized frontier-scale distribution, but not a complete Copilot-specific footprint.

Source

Perplexity

Current product

—

Insufficient

Aurxo did not identify enough comparable public product-level evidence to issue an impact grade in this review.

Le Chat

400-token lifecycle response

—

Moderate

Mistral discloses 1.14 gCO₂e and 45 mL under a broad lifecycle boundary, but not a separate per-response energy value. The overall grade is therefore withheld.

Source

DeepSeek

Current product route

—

Insufficient

Independent references show large differences between DeepSeek-hosted and Azure-hosted scenarios. The current product route must be known before a universal product grade can be issued.

Scoped reference labels

These are valid only for the named historical model, prompt size and deployment assumptions. They are not current universal product grades.

Reference model / scope

Label

Evidence

Metrics and limitation

GPT-4o · March 2025

100 input / 300 output tokens

ChatGPT reference

Independent model

0.423 Wh · 0.148 gCO₂e · 1.953 mL. Not telemetry for every current ChatGPT route.

Source

Claude 3.7 Sonnet

100 input / 300 output tokens

Claude reference

Independent model

0.950 Wh · 0.273 gCO₂e · 5.005 mL. Historical model reference, not telemetry for current Claude routes.

Source

DeepSeek V3

DeepSeek-hosted · 100 input / 300 output

Independent model

2.777 Wh · 1.666 gCO₂e · 19.33 mL. A different hosting route requires a different label.

Source

DeepSeek R1

DeepSeek-hosted · 100 input / 300 output

Independent model

19.251 Wh · 11.551 gCO₂e · 134.004 mL. Reasoning depth and deployment materially affect the result.

Source