AI Agent Context Window Budgeter
Plan how system prompts, few-shot examples, RAG chunks, history, and output reserve consume your model's context window.
Context Budget Inputs
Token counts are planning estimates. Actual tokenizer counts vary by model family. Last updated: 2026-07-07. See notes.
Budget Result
Component Breakdown
| Component | Tokens | % of context | Note |
|---|
Notes
Context-window sizes are the advertised input (or total) token limits for each model family. Actual usable context is smaller once you reserve space for output, safety buffers, and per-request overhead. Use this calculator to plan your agent's context budget before you hit truncation.
- Provider max-output limits may cap the output reserve even when the total context window is larger.
- Tool definitions, JSON schemas, and XML formatting add hidden tokens not captured here — increase the safety buffer for complex agents.
- For multi-turn chats, older turns are often summarized or dropped once the budget is exceeded.
Context Window Cheat Sheet
Compare context limits across models.
Agent Memory / RAG Cost Calc
Estimate embedding and vector DB costs.
Workflow Cost Calculator
Price multi-step agent workflows.
Prompt Tokenizer
Count tokens in a real prompt.
Frequently Asked Questions
Why do I need to budget my context window?
Even models with 128k+ token limits run out faster than expected. System prompts, few-shot examples, retrieved documents, and prior conversation turns all compete for the same space. Budgeting ahead of time prevents silent truncation, degraded answers, and surprise API costs.
What is a 'safety buffer'?
A safety buffer reserves extra tokens for tokenizer differences, formatting overhead, and unexpected prompt growth. We default to 10%; raise it if you use heavy JSON schemas, XML tags, or long tool definitions.
How is output reserve calculated?
The output reserve is a percentage of the total context window set aside for the model's response. Some providers also enforce a hard max_output limit (e.g., Claude 3.5 Sonnet = 8,192 tokens). The calculator uses the smaller of the two values.
Can I use this for local LLMs?
Yes. Select 'Local RTX 4090 (Q4)' or compare against cloud models. Local loaders often support shorter effective context than the theoretical model limit, so treat the result as a practical planning ceiling.
What should I do if I'm over budget?
Shrink few-shot examples, reduce retrieved chunks or chunk size, summarize older conversation history, use a model with a larger context window, or move non-essential instructions out of the system prompt into a separate retrieval step.