# Usage Limits & Balance To ensure fair access for all university members, uniGPT uses a **token-based balance system**. For a comparison of available internal and external models, see [Available Models](). ## How It Works - Every user receives a token balance. - Each prompt and response consumes tokens from your balance. - Your balance **resets every 4 hours**. ## Limits by User Group - **Employees**: receive a balance of **5,000,000** tokens per 4-hour cycle. - **Students**: due to rising provider costs, students now receive roughly **30% fewer credits** per session than before. Employees keep the current limits. > [!TIP] > > To make your credits last longer, prefer cheaper models such as **Gemini Flash** or **GPT 5.6 Luna**, or switch to internal models like **gemma-4-31B-it** or **DeepSeek V4 Flash**. ## Checking Your Balance Click on your **username** in the chat interface to see your current balance. ## Tips for Saving Tokens 1. **Use internal models**: They consume fewer resources and cost the university less (e.g., **gemma-4-31B-it**, **DeepSeek V4 Flash**, **medgemma-1.5-4b-it**, **Qwen3.5-35B-A3B**). 2. **Avoid expensive models for simple tasks**: Use gemma-4 instead of Claude Opus for basic questions. 3. **Prefer Gemini** if you need an external model — it is more cost-effective than Claude Opus or GPT-5. 4. **Lower thinking level**: For supported models, select a lower "thinking" setting to reduce token consumption. ## Related Information - See [Available Models]() for model descriptions, privacy guidance, and relative input/output credit costs. ## Balance Not Resetting? If your balance appears stuck at zero after 4 hours, try **logging out and logging back in**. This usually resolves the issue. ## API Limits The API has limits that are **separate from the chat balance** and are managed independently. By default each API key allows **4 parallel requests**. You can inspect your key's quotas via the `/key/info` endpoint (see [API Access]()). > [!NOTE] > > **Limits are evolving** > > The exact limits are currently being fine-tuned. If you experience issues, please report them in the support channel.