Live router
Backend

All prompts go to this backend. The backend holds the API key; this page never sees it.

OUTPUT-AWARE GREEN INFERENCE ROUTER

Lower-carbon AI chat

Shorter prompts, right-sized answers, and zero-model local resolution for less energy per conversation.

01 Zero-Model Gate 02 PromptShrink 03 Semantic Cache 04 Brevity Cascade 05 CO₂e Receipt

Enter to send · try math, units, hex/RGB, or any question

800×
output vs. input token energy
Ruf & Detyniecki (HotCarbon ’26)
32%
fewer input tokens
PromptShrink v1 corpus benchmark
0 Wh
on local & cached queries
Zero-model gate + session cache
3
right-sized brevity tiers
Sprout · Canopy · Forest
Carbon & Energy Telemetry
Provider-reported token accounting · 0.6 mWh/out-tok · 400 gCO₂e/kWh
Model gpt-4o-mini Cascade ready Zero-Model Gate active (math · units · base · color) Semantic Cache enabled
LIVE
provider-reported
0live answers0 local · 0 cached
0input tokens0 trimmed by PromptShrink
0output tokens
0 Whenergy (est.)0 mL cooling water
0 gCO₂e (est.)0 g at 30 g/kWh (FI)
0cache hits0 Wh reuse
0escalationscascade checks
—saved vs baseline (measured)enable Measure baseline to test
Output Token Footprint vs. Standard Verbose Prompt Awaiting first turn
Arboris right-sized output0 tok · 0 Wh
Standard verbose reference (modeled)0 tok · 0 Wh
Tangible Session Equivalents Estimated
0.0 sof a 9W LED bulb
0 mLcooling water (0.1–0.3 L/kWh)
0 Whmodeled energy avoided
−92.5%if run on 30 g/kWh FI grid
Annual Power & Carbon Savings Simulator (Arboris vs. Unconstrained 220-tok Replies) 50 queries / day
Assumes ~58% average output reduction + 15% zero-model/cache hit rate
1.63 kWh/yrelectricity avoided
653 g CO₂e/yrat 400 g/kWh grid factor
326 mL/yrdata-centre cooling water
181 hrs/yrof a 9W LED bulb

Six levers that cut power vs. standard ChatGPT prompts

Because generating an output token uses ~800× more energy than reading an input token, Arboris attacks both prompt bloat and unnecessary output decoding before, during, and after inference.

1. Zero-Model Local Gate−100% Wh
Intercepts pure arithmetic (12 + 7), percentages (15% of 240), unit conversions (100 km to miles, 60 mph to km/h, 3600 s to hours), base conversions (255 in hex), and colors (#3dd4c0 to rgb) in under 1 ms — 0 GPU calls, 0 tokens, and 0 Wh.
2. Right-Sized Brevity & Fluff Strip−45–70% out-tok
Standard ChatGPT replies add polite preambles ("Certainly! Great question!") and recaps. Arboris strips AI pleasantries, routes lookups to Sprout (~2 sentences), explanations to Canopy (~4), and deep queries to Forest (~6), with expand-on-demand.
3. PromptShrink & In-Place Compress−32% in-tok
Strips conversational filler, email sign-offs, and wordy clauses before sending — or click ✂ Compress in the composer to shrink your prompt in place — while preserving code fences and instructions.
4. Exact & Semantic Session Caching0 new tokens
Reuses verified high-quality answers across identical or lexically equivalent fact/definition queries in your session — eliminating redundant GPU prefill and decode passes completely.
5. Reasoning Effort & Small-Tier CascadeCapped CoT
Caps hidden chain-of-thought reasoning_effort on trivial queries and routes easy lookups to a smaller model tier first, escalating to the primary model only when confidence or quality score is low.
6. Output-Weighted Carbon ReceiptsTransparent
Every answer includes an interactive energy, CO₂e, and cooling-water receipt computed from provider-reported tokens (including hidden reasoning_tokens) with published academic anchors.

Routing & brevity modes

In Auto mode, Arboris selects a sentence budget and token cap before calling the model:

  • Sprout (≤12 words): ~2 sentences, 120 max tokens
  • Canopy (13–34 words): ~4 sentences, 240 max tokens
  • Forest (35+ words, complexity marker, or 2+ questions): ~6 sentences, 400 max tokens

Need a fuller explanation? Click More detail under any live answer to request an expanded Forest response.

How the numbers work

Generating tokens uses far more compute than reading prompt tokens:

  • Energy: output × 0.0006 Wh + input × 0.0006 / 800 Wh (Epoch AI 2025; Ruf & Detyniecki HotCarbon ’26)
  • Carbon: Wh / 1000 × 400 gCO₂e/kWh (0.00024 g per output token)
  • Cooling water: Wh × 0.1–0.3 mL (closed-loop data-centre WUE; generation water excluded)

All figures are arithmetic estimates from provider-reported tokens, never physical data-centre measurements.

Finland reference & limits

Clean-energy reference: Google Cloud Finland (europe-north1): ~98% carbon-free energy, 30 gCO₂eq/kWh.

Honest note: the hosted backend runs in us-central1, and OpenRouter does not disclose where upstream model inference runs. Live receipts therefore use the 400 g/kWh illustrative grid factor.

Google Cloud region carbon data →

How Arboris works

Live only

Live answers on this page come from a real model through the backend: on the hosted site that backend is already running, and on your own machine it is python server/app.py. Every question goes to a real model through the backend. There is no offline fallback: if the backend is unavailable, check the backend URL in settings.

What the steps really do

The numbers are estimates

Privacy

This page loads no outside fonts or third-party scripts and uses no analytics. Your questions go to the backend you configured, and from there to its model provider, whose data-retention policy applies. Your chat is saved only in this browser's localStorage under arboris-live-v1; the ✕ Clear button deletes it.