Token Forecast

Know what a message will cost before you send it.

A free browser extension for claude.ai that predicts how much of your 5-hour and weekly Claude limit a message will use, tells you which model and effort level are enough (with reasons), and learns your plan over time. Everything runs in your browser.

⬇ Add to Chrome — free, from the Chrome Web Store

Chrome Web Store · Website · Releases · Türkçe README

▶ Watch the full walkthrough on YouTube (4 min) · Türkçe

47-second intro video and a live demo of the real prediction engine are on the website. The interface in the video is illustrative; the numbers come from the extension's own engine.

---

Why this exists

Claude Code shows your usage. The claude.ai chat screen does not. On claude.ai you open Settings → Usage, and there you only see what you have already spent. Limits often run out in the middle of a job, on a message you did not expect to be expensive.

Token Forecast brings that into the chat and goes one step further: it tells you before you press send.

Feature	Settings → Usage	Usage meters*	Token Forecast
5-hour and weekly percentage, reset time	✅	✅	✅
Inside the chat, next to the message box	❌	✅	✅
What this message will cost, before sending	❌	❌	✅
Which model is enough, and why	❌	❌	✅
Hidden cost of switching model mid-chat	❌	❌	✅
Roughly how many tasks the remaining limit allows	❌	partly	✅
Learns from your plan and habits	❌	❌	✅
"Why was that message so expensive?" explanation	❌	❌	✅
Weekly report, hourly heatmap, personal tips	❌	partly	✅
Cache and peak-hour countdowns	❌	partly	✅

* We looked at the pages of eight usage extensions in October 2026. None offered a per-message estimate before sending or a reasoned model suggestion; all show usage after sending. One has a cache countdown, one a peak-hour countdown, one a heatmap.

Screenshots

A forecast and a reasoned suggestion above the message box

A small pill shows the percentage of your limit the message will use, with an 80% range. Open Why? to see the detected task, the topic, what the extension looked at, and the basis for the suggestion (Anthropic's model-choice guide). If it is not sure, it never suggests a light model.

Effort suggestions

From Low to Max: which effort level is enough for this task, and how much limit you would save if your current setting is higher. It knows each model's default and the Extra high level.

It looks at the task, not the word

"I need a website with a translation system" is a build job, not a translation. The extension tells the difference and does not suggest Haiku. The automated tests contain 15 trap sentences like this plus real translation requests.

Toolbar window

Your 5-hour and weekly limits, what the remaining limit still allows (short questions, explanations, code fixes…), the weekly report, and the state of your personal calibration.

Weekly report

Messages, limit spent, peak-hour share, model and task mix, topic and effort mix, a 7×24 hourly heatmap, the most expensive messages, forecast accuracy, a learning-proof chart and personal tips. Computed only in your browser from text-free records. Monday-morning notification, CSV export.

---

What it does

Before you send

• Forecast before sending. Prompt, attachments (images, PDF, Word, Excel, text), chat history, model, effort and cache state → percentage of your limit, with an 80% range.
• Reasoned model suggestion. Which model is enough, what it looked at, the basis, and how much you would save. 👍/👎 and a "wrong task type?" correction tune it to you.
• Topic recognition. Classifies a message into 16 topic areas and sub-topics (Science · Physics, Engineering · Electrical…) using about 2,600 keywords. For health, law and money questions it never suggests a light model.
• Effort suggestion. Low to Max, with the savings if your setting is higher. Creative writing is capped at Medium; proofs get at least High; hard coding on Opus can reach Extra high.
• Cost breakdown. How much of the cost comes from the answer, the reasoning, the history and the attachments.
• Cache and peak-hour countdowns. How many more minutes the chat history is read for free, and when peak hours end.
• A pill you can move. Drag it anywhere (or use Alt + arrow keys). It steps aside by itself when claude.ai's own model or effort menu opens.

While you spend

• At this pace, when does it run out? Looks at your last hour and tells you whether the 5-hour limit fills before it resets.
• Hand-off summary. When a chat gets too long, one click copies a ready prompt asking Claude for a short summary to carry into a new chat.
• Warnings and notifications. "May not fit", context window, cache and peak hours; 50% / 75% / 90% thresholds and a "limit reset" notification; a toolbar badge.

After you send

• "Why was that so expensive?" When the measured usage is far above the forecast, it lists likely reasons: history re-read after a long pause, peak hours, a long answer, or use on another device.
• Weekly report (see above).

This is the beginning of the README. Read the rest at the original. Read the full README ↗