DEV Community

Cover image for Running GPT-6.2 Astra or Gemini 4 Argon to format a JSON file? I built a local tracker to see what that actually costs
IB2M_Official
IB2M_Official

Posted on

Running GPT-6.2 Astra or Gemini 4 Argon to format a JSON file? I built a local tracker to see what that actually costs

With GPT-6.2 Astra, Gemini 4 Argon, and Claude Opus 5.5 / Fable 5.2 sitting right in the default model dropdowns, it has never been easier to throw massive reasoning compute at completely trivial tasks.

Earlier this week, I caught myself leaving a flagship reasoning model active while asking it to fix a 10-line CSS grid and rename a few variables. A lightweight model like GPT-6.2 Sol, Gemini 4 Hydrogen (or Flash-Lite), or Claude Haiku 5.5 handles that in milliseconds for a fraction of the power. Because web chat interfaces hide everything behind a clean loading spinner, there is zero feedback loop telling you how heavy a prompt actually is.

I built EcoPrompt to fix that blind spot without turning into a preachy dashboard or harvesting user data. It is a free, open-source Chrome extension that estimates the physical and financial footprint of your prompts in real time, 100% locally in your browser.


How it looks in practice

Whenever you hit enter on ChatGPT, Google Gemini, or Claude, EcoPrompt injects a small, unobtrusive toast in the corner of the page showing four numbers for that specific interaction:

  • Water (mL): Direct datacenter evaporative cooling + indirect water from grid electricity.
  • Energy (Wh): Active compute draw across prefill and decode phases.
  • Carbon (g CO₂e): Estimated emissions calibrated on datacenter PUE and regional grid averages.
  • API Cost Equivalence: What that exact prompt would roughly cost at public token rates.

EcoPrompt in-page toast on Gemini

Clicking the extension icon opens a compact daily summary so you can see where your usage is going at a glance:

EcoPrompt popup on Claude


Different models, very different footprints

Not all prompts hit hardware the same way. EcoPrompt detects which platform and model tier you are using in the UI to adjust its compute estimates accordingly:

  • Fast & lightweight tiers (Gemini 4 Hydrogen, Flash-Lite, Claude Haiku 5.5, mini models): Built for speed and low token overhead (~0.3 Wh per standard query).
  • Everyday frontier workhorses (GPT-6.2 Sol, Claude Sonnet 5.5): Balanced compute for daily coding and writing tasks.
  • Heavy reasoning & agentic tiers (GPT-6.2 Astra, Gemini 4 Argon / Carbon, Claude Opus 5.5 & Fable 5.2): Long hidden chains of thought and sustained KV-cache residency multiply GPU cluster time and water cooling requirements significantly.

Under the hood (and why there is zero telemetry)

An extension running on chatgpt.com, gemini.google.com, and claude.ai has access to sensitive workspaces, so privacy had to be non-negotiable from day one:

  1. Zero prompt logging: EcoPrompt never reads, stores, or transmits your prompt text. Character length is evaluated ephemerally in memory just to weight prefill/decode pressure, then immediately discarded.
  2. No external servers: There is no backend, no analytics script, and no account required. Everything stays in chrome.storage.local.
  3. Pure Vanilla JS (< 150 KB): Built on Manifest V3 with zero npm dependencies. The analytics timeline in the full dashboard runs on a custom native SVG chart renderer (~8 KB) instead of bundling Chart.js or D3.

EcoPrompt full analytics dashboard

I also added a soft Daily Budget & Quotas tab in the dashboard. You can set a daily prompt, energy, water, or CO₂ target—not to block your work, just to nudge yourself into batching prompts or switching back to Haiku / Flash / Sol when you don't need heavy reasoning.

EcoPrompt daily quotas


Try it or check out the code

EcoPrompt is completely free and open-source (CC BY-NC-SA 4.0). If you want to test it out, audit the calculation heuristics, or open a PR to support another provider:

Curious how other devs here manage their model switching—do you actively drop down to lighter models like Haiku 5.5, Sol, or Hydrogen for quick tasks, or just leave your flagship model pinned all day?

Top comments (1)

Collapse
 
ib2mofficial profile image
IB2M_Official •

Quick note for anyone inspecting the repo: the provider DOM selectors and model tier weights live in a clean config module, so adding support for interfaces like Perplexity, Mistral, or DeepSeek is pretty straightforward.

If there's a specific AI web interface or local setup you use daily and would like tracked, drop it below and I'll prioritize it for the next release!