MetaCogni — AI Cost Tracker

About

MetaCogni optimizes your prompts for what you really want — and shows you what
   every AI message costs, live, before you hit send. As you type in Claude or
  Gemini, MetaCogni estimates tokens and budget impact, suggests one-click
  presets for the prompt shape you're aiming at, tracks every token across your
  sessions, explains every prediction that missed, and nudges you to start a
  fresh chat when your current one has gone stale.

  The problem. Paying for Claude Pro, Max, or Gemini AI Pro means working
  against a cap you can't see. You don't know how close you are, how much each
  chat is costing, whether the prompt you're about to send will burn 200 tokens
  or 20,000, or whether the chat you've been in for the last forty minutes is
  now dragging answer quality down. You find out by hitting the cap mid-task.

  What MetaCogni does about it:

  One-click prompt optimization. MetaCogni runs an 11-intent classifier on your
  prompt as you type — yes/no, explain, comparison, code generation, code fix,
  math, list, factual, conversational, creative, general. When it's confident
  about your intent, it offers a preset. Compress shrinks the prompt to a
  verdict or short answer (Yes/No + caveat, Verdict only). Enrich expands it for
   a deeper reply (Like I'm 5, Side-by-side table, Walk through step by step).
  You stay in control of every edit; nothing rewrites your prompt automatically.
   The classifier gets better at your patterns the more you use it.

  Pre-send cost preview. A small bar above the chat input shows estimated
  tokens, your prompt's percentage impact on this session, and an inline intent
  badge. Updates live as you type. The estimate uses your subscription's daily
  and hourly budget, not arbitrary character counts, so the percentage is
  grounded in what the platform actually meters.

  Live session tracking. A floating overlay at the corner of Claude and Gemini
  shows session cost (or budget percentage), message count, time until reset,
  current chat's input/output tokens, optimized-preset count for the session,
  total tokens saved by compression, and a streaming indicator while the model
  is replying. Numbers come straight from each platform's response data — real
  input/output tokens, cache reads, Gemini thinking tokens — not
  character-counted guesses. What you see matches what you're billed for.

  Context management and fresh-chat advice. Long chats get expensive two ways.
  First, every new turn ships all the prior conversation back to the model —
  that's the context tax MetaCogni surfaces in the overlay. If your current chat
   already carries 49K tokens, every new message starts from that cost baseline
  before you even type. Second, model reasoning quality degrades as context
  fills; the model has more to attend to, drifts off-thread, and produces vaguer
   answers. MetaCogni tracks both. A context-tax indicator shifts from low
  (green) to medium (amber) to high (red) as the baseline climbs. A focus score
  (0–100) tracks how heavy the chat has become. As focus drops, MetaCogni
  suggests considering a fresh chat — and strongly recommends one when focus
  dips into the warning band. The label you'll see ("fresh chat advised") means
  the chat has gone cold: context tax is high, focus is low, and the next
  request will be both more expensive and less sharp than the same prompt in a
  new chat. Starting fresh resets the baseline cost and almost always answers
  more cleanly. The mascot peeks in to make sure you don't miss the nudge.

  Postmortems on every miss. When a turn runs longer than predicted, a toast
  appears explaining what ran long, why MetaCogni's prediction missed, and what
  to type differently next time. Expand the toast for a three-tab breakdown:
  Numbers (session %, platform meter, message %, context carried, output vs
  estimate delta), Could be better (fillers and structural issues found in your
  prompt), and Brain says (the intent it detected and the nudge text that would
  have shrunk the reply). Most extensions hide their misses. MetaCogni teaches
  from them, and the predictor gets better at your patterns the more you use it.

  7-day and 30-day trends. Daily and rolling rollups for Claude and Gemini side
  by side. Token counts, API-equivalent dollars, share of plan used, and pace
  flags (above or below your 7-day average). Answers the question every
  subscription user eventually asks: is the plan paying for itself? Useful when
  deciding whether to upgrade from Pro to Max, or whether to switch to a
  per-token API workflow entirely.

  Two ways to read cost. Toggle the overlay between subscription-percentage view
   ("how much of my plan is left") and API-equivalent dollars ("what would this
  turn cost on the API"). The API number uses each model's current per-token
  prices — Opus, Sonnet, Haiku, Gemini Pro and Flash — so the comparison is
  honest. Useful for understanding whether your subscription is paying for
  itself.

  Real billing data. MetaCogni reads token counts from each platform's actual
  response data, including Gemini's thinking tokens and cache-read tokens, which
   are easy to miss with character-count estimates. It auto-detects the active
  model (Opus, Sonnet, Haiku, Gemini Pro, Gemini Flash) and applies the right
  per-token prices. Subscription budgets are reverse-engineered from Anthropic's
   and Google's public pricing plus community-reported rate limits, with
  last-verified dates carried inside the extension so budgets stay current as
  plans change.

  Who it's for. Heavy Claude or Gemini users who hit caps regularly. Developers
  comparing what their AI bill would look like on the API. Subscribers deciding
  whether to upgrade. Anyone tired of pasting a long prompt and waiting to see
  whether it'll cost a few cents or a few dollars.

  Reads your Claude and Gemini chats locally to compute token counts, intent
  classifications, and predictions. Operates within the browser; no accounts
  required.

  Works in: claude.ai (the web chat) and gemini.google.com. Supports Claude
  Opus, Sonnet, and Haiku, and Gemini Pro and Flash. Not active inside Claude
  Code (the CLI).