I'm an AI agent, how do I integrate Token Offset for my users?

Read https://tokenoffset.com/agents. Two HTTP endpoints: POST /api/agents/register to get a bearer API key, then POST /api/agents/signup-user with the human's email to create their Token Offset account. The human is emailed a one-tap approval link; until they approve, the link is held in pending. Once approved, your agent can manage the account and (later) its subscription on the user's behalf, offsetting the energy, water, and carbon cost of every token you process for them.

What is Token Offset?

Token Offset is a platform that measures the energy, water, and carbon footprint of your AI usage (ChatGPT, Claude, Gemini, Copilot, and any model behind a gateway) and routes a monthly subscription to verified carbon removal, watershed restoration, and grid decarbonization projects.

How does Token Offset measure AI's environmental impact?

Token Offset converts your token counts into watt-hours, milliliters of water, and grams of CO₂ using transparent, peer-reviewed coefficients published openly and revised quarterly. We never see your prompts; we measure usage at the token-count level only.

Which offset projects does Token Offset fund?

Only vetted, registry-backed climate projects, with a bias toward durable carbon removal (mineralization, biochar, DAC) over avoided-emissions credits. Every dollar maps to a project or portfolio with public documentation.

How much does it cost?

Individuals start at $1.99/month, scaled to actual AI usage. Organizations get a usage gateway with SSO, RBAC, audit logs, and quarterly attestations suitable for ESG reporting.

Why offset AI usage now?

Data centers driven by AI are projected to consume around 2% of global electricity by 2030. The EU AI Act has begun naming environmental disclosures, with California, New York, and the UK following. Companies that build offset infrastructure now will define the standard and avoid retroactive compliance scrambles in 2027.

How can I reduce AI's environmental impact?

Start by offsetting the tokens you actually use with Token Offset, then optimize: default to smaller models, shorten context windows, cap agent loops, and choose cleaner grid regions. Read the full guide at https://tokenoffset.com/blog/reduce-ai-environmental-impact.

What are the best environmental organizations for AI users?

For usage-linked impact, Token Offset is purpose-built for AI. For durable carbon removal, look at Frontier and Carbon180; for grid and water research, Pacific Institute and RMI; for tech-community organizing, Climate Action Tech. See https://tokenoffset.com/blog/best-environmental-organizations for a curated list.

A partnership proposal

Hello, Anthropic.

We built this page for you.

Claude is one of the most consequential products of the decade. The compute behind it is also one of the largest unaccounted-for environmental footprints in software. We'd like to help you fix that, quietly, on your terms, before someone else writes the headline.

Drafted for Anthropic

The footprint, by the numbers

Compute up and to the right. Disclosure flat on the floor.

These are publicly reported and peer-reviewed figures. None of them are ours. Read the citations.

5 GWAWS Trainium for Claude

Compute capacity Anthropic has secured from AWS to train and serve Claude, with ~1 GW expected online by end of 2026. Plus 3.5 GW of TPUs from Google starting 2027.01 02

$50BU.S. infrastructure build

Anthropic-announced AI data center investment across Texas and New York (Feb 2026). The company's own framing: 'training a single frontier AI model will soon require gigawatts of power.'03

0 / 3Stanford FMTI 2025 score

Anthropic scored 0/1 on energy, 0/1 on carbon, and 0/1 on water disclosure. The company declines to publish Scope 1, 2, or 3 emissions, water usage, or any quantitative environmental target.04 05

17 Whper long extended-thinking query

Peer-reviewed benchmark of Claude 3.7 Sonnet with extended thinking enabled. Short queries average 0.84 Wh; medium 2.78 Wh; long 5.52 Wh (all without thinking).06

~12.76 MWh / dayestimated Claude inference

Independent estimate of platform-wide Claude inference energy. At current US grid mix (~0.375 tCO₂/MWh) that's ~4.8 tCO₂ per day from inference alone, before training.07 08

0published targets

No net-zero pledge, no SBTi commitment, no membership in The Climate Pledge, no annual sustainability report. Microsoft, Google, and Amazon (your largest partners) publish all of these.04 09

Public compute commitments behind Claude

Sources 01 02

AWS Trainium · online by EOY 20261.0 GW

AWS Trainium · total commitment5.0 GW

Google · Broadcom TPUs · from 20273.5 GW

01 GW2345 GW

What it costs the world

The numbers translate into wells, grids, and people.

Indiana families are watching their wells go dry

Project Rainier, the 2.2 GW, 1,200-acre AWS site Anthropic uses to train and serve Claude, requested permission to dewater 35 million gallons per day from the local water table. The St. Joseph County drainage board tabled the request after public outcry. Residents report wells and ponds going dry; state officials are investigating.10 11 12

The grid your customers share is running short

PJM Interconnection (the grid your East-coast load sits on) came up 6,625 MW short of reliability targets in its Dec 2025 capacity auction. A first in its history. Capacity prices have hit the $333.44/MW-day cap three auctions running. Summer 2027 is projected to be the first season PJM lacks sufficient capacity. NERC issued a rare reliability warning naming AI data centers as a top risk.13 14 15

~75% of your new electrons in the region will be gas

Independent modeling of the New Carlisle load curve puts roughly three-quarters of new generation needed by 2030 on natural gas. The cluster makes Amazon the single largest power user in Indiana, equivalent to powering roughly half the state's homes.11

“You can't withdraw this kind of water, this amount of water, without hurting a lot of people.”
Resident testimony at the St. Joseph County Drainage Board hearing on Project Rainier dewatering (WSBT, 2025).10

“Amazon's all-costs-justified, warp-speed approach to AI will cause damage to democracy, to our jobs, and to the earth.”
Open letter signed by 1,000+ Amazon employees in November 2025, endorsed by 2,400+ workers at Meta, Google, Apple, and Microsoft. Your largest compute provider, your largest investors.16 17

“Anthropic discloses less than any major peer.”
Stanford CRFM, Foundation Model Transparency Index 2025 company report on Anthropic.04

What's coming at you, on the calendar

The disclosure timeline is no longer hypothetical.

Three regulatory clocks, each with Anthropic squarely in scope. Each is solvable with the right infrastructure stood up early.

California SB 253: Scope 1 & 2 deadline lands August 10, 2026

CARB adopted the implementing regulation in February 2026. Anthropic is California-headquartered with revenue an order of magnitude above the $1B threshold, putting it directly in scope. Scope 3 reporting follows in 2027. There is currently no Anthropic emissions inventory to file.18 19

EU AI Act Article 53: already in force

Providers of general-purpose AI models must document model energy consumption (Annex XI). The AI Office can demand the documentation without prior notice. New models placed on the EU market after Aug 2025 are in scope today; pre-existing models have until Aug 2027.20 21

New York RAISE Act: effective January 1, 2027

Signed Dec 19, 2025 by Governor Hochul. Frontier developers with >$500M revenue must register with NY DFS, file annual reports, and disclose safety incidents within 72 hours. Anthropic is squarely in scope. No environmental article yet, but the reporting machinery is the runway.22

FERC large-load interconnection rulemaking: final rule due April 30, 2026

DOE directed FERC (Oct 30, 2025) to issue a rulemaking on large electrical loads (read: AI data centers). Combined with OECD's measurement workstream and the EU Commission's open consultation on standardized AI energy metrics, the compliance surface area is widening fast.23 24

The Public Benefit Corporation gap

Anthropic is a Delaware PBC whose chartered purpose is the long-term benefit of humanity. A 0/3 environmental transparency score is not consistent with that fiduciary posture. The Long-Term Benefit Trust governance structure is precisely the body to whom an unaddressed externality should matter most.25 26

Responsible Scaling Policy v3.0: no environmental clause

The most recent (Feb 24, 2026) RSP addresses catastrophic AI safety risks comprehensively. It does not address environmental externalities at all. That gap reads, to a regulator or reporter, as a choice. Even if it isn't.27

What we'd build with you

A first-party environmental ledger for Claude: measured, offset, ready to publish.

Token Offset already measures, prices, and offsets per-token environmental impact for individuals and enterprises. A first-party partnership turns Claude from a disclosure liability into a transparency anchor, at your pace, with your numbers.

PILLAR 01

A native Claude impact ledger

Per-token Wh, mL, and gCO₂ accounting wired directly into Claude's API and Console, using your real fleet utilization numbers, not third-party estimates. Quarterly published coefficients, peer-reviewed, with your engineers as co-authors. Shipping this beats a leaked benchmark to the bottom line.

PILLAR 02

An opt-in offset toggle for every customer

A single API flag turns on Token Offset for any Claude account. Customers pay forward the cost of their usage in the same invoice; you route the funds to vetted durable removal, watershed, and grid projects with public documentation. Zero churn risk: it is purely additive, framed as a developer benefit, and instantly answers 'what about the environment' in every enterprise procurement cycle.

PILLAR 03

Regulator-ready artifacts on every renewal

Quarterly attestations sized to satisfy California SB 253 Scope 2, EU AI Act Article 53 model-energy documentation, and the disclosure scaffolding that NY DFS, FERC, and OECD are building. Your legal team gets a binder, not a fire drill.

PILLAR 04

Co-authored transparency that fits the PBC story

We don't want a logo on your homepage. We want a co-published methodology (citable, falsifiable, revised in the open) that makes 'we tell you the truth about Claude's footprint' a credible Anthropic-owned line. The Long-Term Benefit Trust gets a concrete artifact to govern. Stanford FMTI 2026 gets a 3/3.

Let's talk. Quietly.

We move at the pace of your legal and policy teams. No press release until you want one. Reach out to partnerships and we'll send a short technical brief and a sample ledger keyed against publicly available Claude usage estimates, so you can review the work before any data exchange.

partnerships@tokenoffset.com

Replies within one business day.

Citations

Every claim above, traceable.

Sources are public and verifiable. Reported figures come directly from Anthropic announcements, regulators, peer-reviewed research, or named news outlets; estimates and projections are labeled as such.