Hitting your token limit can interrupt important work at inopportune moments. The recommendations below detail how you can decrease your token consumption. They apply to the approved AI tools available in the GP AI Portal.
A token is the basic unit of text processed by LLMs and is slightly less than a typical word in length. Both the information that you provide to an LLM and the model’s generated response consume tokens.
The context window is the model's working memory for a given conversation. The LLM’s response quality can degrade as excess information is entered into the context window. Limiting the amount of unnecessary information in the context window can save tokens and improve response quality.
Match the model to the task
Higher-end models generate longer, more detailed responses by default, consuming more tokens. For routine tasks, such as reformatting a document or drafting a standard email, a lighter model such as Claude Haiku or ChatGPT GPT-5.5 can produce output that is just as usable at a far lower token cost. Reserve Sonnet, Opus, and GPT-5.6 for tasks that require deeper reasoning. Select the model from the model menu when starting a new chat.
Convert PDFs and Word documents to Markdown before uploading
LLMs typically consume far more tokens to process a PDF or Word document than to process a Markdown file with the same content. Converting your PDF or Word documents to Markdown files before uploading the text into Claude is a valuable method to lower your token usage. Convert your chosen document using the PDF-to-Markdown Converter and the Word-to-Markdown Converter on the GP AI Portal’s Tools for AI page, and upload the resulting file into your LLM of choice.
Upload only the sections relevant to your query
Uploading a full document when your question concerns one section uses unnecessary tokens and can reduce response quality. Extract and paste the specific sections or paragraphs you need. If working across several parts of a document, reference the relevant page or paragraph numbers in your prompt rather than uploading the whole document.
Start a new conversation rather than extend a long chat
Every message causes the model to reprocess the entire conversation history. When a conversation has run beyond roughly ten exchanges, ask the model to produce a handover summary capturing the key conclusions, open questions, and essential context. Paste the summary into a new chat and continue from there.
Use Projects when working with multiple documents over time
The Projects feature uses Retrieval-Augmented Generation (RAG), a process in which the model retrieves only the portions of your document most relevant to your query. The Projects feature is particularly useful for matters that you plan to work on for an extended period of time. Create a Project for the matter, upload the core documents as context within the Project, and work from the Project rather than re-attach files to each new conversation.
Write specific prompts that call for concision
A prompt that states the task clearly, identifies the relevant document, and specifies the format and length of the desired output will produce a more concise answer at lower cost than a vague prompt.
Disable tools and connectors that you are not using
Connections for web search, file storage, and third-party integrations contribute to token usage when enabled, even when they are not actively called during a conversation. Toggle off unused tools in the conversation’s settings. Enable these tools only when the task requires real-time information or data from a connected system.