Token Offset
Documentation

The math, with receipts.

Every estimate on the dashboard derives from three impact constants and one contribution rate, applied to an estimated monthly token count, either the band you picked (Casual / Worker / Agentic / Extreme) or a 90-day sample from a connected admin key. The constants are conservative midpoints calibrated against peer-reviewed studies, production telemetry, and operator-published lifecycle assessments; the rate is set to fund high-quality durable removal at a price normal people might still choose to pay. This page is the live record of how those four numbers are chosen, and why.

Current algorithm

Last revised

For any token count n, we compute four numbers using the constants below. The contribution amount is rounded up to the nearest cent so the payable price never under-collects the published rate.

Energy
n × 0.0005 Wh
0.50 Wh per 1k tokens
Water
n × 0.003 mL
3 mL per 1k tokens
CO₂e
n × 0.0006 g
0.6 g per 1k tokens
Contribution
ceil cents((n / 1000) × $0.00025)
$0.25 per 1M tokens

Conversion · 1,000 tokens ≈ 750 words

Try the numbers

Pick a monthly token volume, then dial in how much of it you want to cover. 100% means the full published contribution for your estimated footprint. Push past 100% if you want to fund removal work for tokens you didn't generate yourself: every percentage point above the line goes to permanent drawdown, watershed restoration, and grid decarbonization.

~3,750,000 words
tokens / mo
100% covered
100% of your footprint

How much AI does anyone actually use?

The four bands below frame the order-of-magnitude differences between an occasional ChatGPT user and a developer leaning on coding agents every day. Each band cites the public study or operator disclosure we calibrated it against; the "Monthly midpoint" column is the single value we feed into the calculator above when you tap that preset.

As more Token Offset users connect admin keys, we additionally tune the midpoints against the aggregated, de-identified 90-day samples that flow through the platform, so each band stays grounded in what people actually use, not just what studies predicted. See what we save and what we do with it for the specifics; identifiable usage is never shared externally.

BandWhat it looks likeMonthly midpointOffset / mo
Casual
Occasional questions to Claude and ChatGPT
A few thousand tokens on use days, or roughly a few dozen short conversations per month.100K$0.03
Worker
Average human knowledge worker, daily AI use
20,000 to 250,000 tokens per workday for writing, research, analysis, and repeatable GPT/project workflows.5M$1.25
Agentic
Coding agents and multi-step workflows
Hundreds of millions of tokens per month for daily coding agents, long-context tasks, retries, and context replay.300M$75.00
Extreme
Internal devs, eval pipelines, agent fleets
Up to 210 billion tokens in a single week at the very top of the distribution.10B$2,500
Casual · OpenAI / NBER (Sept 2025) reports ~18B ChatGPT messages/week from ~700M weekly users (~26 messages per weekly active user). We pair that with OpenAI's 1 token ~= 0.75 words rule and keep this band below the all-user average to represent occasional users.
Worker · OpenAI 'State of Enterprise AI' (2025): 7M+ workplace seats, ~8x Enterprise message growth, ~30% more messages per worker, 19x growth in Projects/GPT workflows, and 320x growth in reasoning-token consumption per organization.
Agentic · IDE-Bench reports ~0.18M-1.35M tokens per successful IDE-agent task, ProjDevBench averages 4.81M tokens/problem for end-to-end project tasks, and Cursor users report 1M-6M token agent requests under heavy use.
Extreme · Cursor public forum disclosures: Cursor is Anthropic's largest API customer and has saturated their GPU capacity at points; internal devs running agent fleets sit at the extreme tail of usage.

Energy per token

We use 0.0005 Wh per token (0.50 Wh per 1,000 tokens) as a blended average across input, cache, and output tokens. It sits near the public prompt-level evidence: Google's production Gemini fleet logs 0.24 Wh for the median text prompt, Epoch AI estimates 0.30 Wh for a typical short GPT-4o query, and Jegham et al. measure 0.42 Wh per short GPT-4o query. Long-context, multimodal, and reasoning workloads can be several Wh higher.

Water per token

We use 0.003 mL per token (3 mL per 1,000 tokens). The number combines on-site cooling water with a conservative slice of the water embedded in electricity generation. The published range is wide on purpose: Google reports 0.26 mLon-site water for a median Gemini prompt, while Mistral's auditor-reviewed LCA reports 45 mL for a 400-token Le Chat response when upstream water is included.

Carbon per token

We use 0.0006 g CO₂e per token (0.6 g per 1,000 tokens). A pure electricity-only calculation from our energy constant times the EPA eGRID 2023 grid factor would be lower; we round up toward lifecycle results like Mistral's because servers, networking, allocation overhead, and amortised training emissions don't disappear just because providers don't publish them per token.

Contribution rate

We suggest $0.00025 per 1,000 tokens, or $0.25 per million tokens. That is not a claim that one thousand tokens cause exactly 0.025 cents of damage. It is a practical high-quality offset contribution: low enough that regular users can opt in, high enough that pooled monthly totals can buy credible durable removal rather than fractional pennies of cheap avoidance credits alone.

The math is intentionally visible. At 0.6 g CO₂e per 1,000 tokens, the full contribution implies roughly $417 per tCO₂e all-in. We target about 70% of that for durable carbon removal, which leaves roughly $292 per tCO₂e for the removal credit itself. That lands near current higher-quality biochar and enhanced-weathering ranges, though below Frontier's full offtake portfolio pricing. The remaining 30% covers watershed restoration, grid work, verification, payments, and public methodology upkeep.

In user-facing terms, the current 100% rate is about $0.03/month for Casual, $1.25/month for Worker, $75/month for Agentic, and $2,500/month for Extreme. The sliding scale lets people choose a lower or higher percentage without changing the underlying allocation quality.

Caveats

  • These are estimates, not measurements.No major provider publishes per-token energy by deployment. We'll refresh the constants as better operator data lands; this page is the live record of what we believe today.
  • Cache tokens count. Cached input still uses memory bandwidth, GPU time, and compute to assemble: and Anthropic still charges for it. We sum all token columns (uncached input, cache read, cache creation 5m + 1h, output) when computing impact and offset.
  • Offsets are not a substitute for reducing use. The cheapest token, environmentally, is the one you never generated. We surface impact at the day level on the dashboard so the feedback loop is short.
See /docs/connections for exactly what we read from each provider, what we don't, and how the encrypted key storage works.