Which AI tools expose real usage limits (and which don't)
A practical map of what you can actually track today across Claude, ChatGPT, Codex, Grok, Perplexity and Gemini.
Not every AI tool tells you how much you have left. Here’s the honest state of play for the major ones.
Full live limits
- Claude - 5-hour + weekly windows, plus per-model buckets. Read as a percentage with exact reset times.
- ChatGPT - per-feature caps (deep research, image generation, file upload) with remaining counts and resets.
- Codex - 5-hour and weekly usage windows, as a percentage.
- Grok - a per-window request quota (remaining / total).
- Perplexity - the Pro-search allowance (remaining).
Activity only
- Gemini - Google exposes no usage/limit endpoint on the free tier, so the best any tracker can do is confirm you’re active and warn about peak hours. Anyone claiming a live Gemini free-tier counter is guessing.
The takeaway
If a tool has a web session, its own app already knows your limit - a tracker just reads the same numbers. Tools without a web session (CLI agents) need an API key. And a few, like Gemini free, simply don’t publish the data.
AI Karma Tracker sticks to real numbers where they exist and is upfront when they don’t.