Every prompt in Claude Cowork against a Project re-reads every uploaded page, for every user. Connect the same Claude to a raia agent over MCP and the agent searches one shared vector store, distills the matches with a smaller model, and hands back only what the question needs. Drag the sliders to see what that does to tokens, cost, and headroom.
Documents live in project knowledge. Each user's prompts carry the whole library into context.
Documents live in one raia knowledge base. Claude calls the agent over MCP; it searches, distills, and returns only what is needed.
Every "Run one prompt" adds that prompt's tokens and cost to each side. Run it a few times and watch the gap open.
Project cost climbs with every page. raia stays flat because the knowledge base, not Claude's context, holds the library, so a prompt reads the same few passages whether the index holds 10 documents or 2,000. Switch to a log scale to see both lines at once.
How to read this. Pricing uses Claude Opus 5 API rates: $5 per million input tokens, $25 per million output, cache reads at 10% of input. raia's distilling step is priced at the selected model's rates (Claude Haiku 4.5 by default, $1 in and $5 out) plus raia's markup on those tokens (50% by default), and is included in the raia cost. raia lets you choose the agent's model, so the picker also carries OpenAI GPT-5.6 Luna ($0.20 in, $1.20 out) and Google Gemini 3.8 Flash ($0.75 / $3.75, list price through 2026) and Gemini 3.5 Flash-Lite ($0.30 / $2.50) at their September 2026 API list prices. The open-weight preset is a placeholder rate; replace it with your hosting provider's price. A Claude project or Cowork seat is billed by subscription, so the project cost shown is what the same token traffic would cost at API rates, which is what the platform must absorb behind the seat.
What is left out. One-time embedding and hosting for the raia knowledge base, the raia subscription itself, cache-write premiums on the first prompt, and thinking tokens on either side. Retrieval quality depends on chunking and embeddings; the signal share is an illustration, not a benchmark. Project capacity defaults to the 1M-token Opus 5 context window. Claude.ai and Cowork may enforce a lower project-knowledge limit; lower the value under Assumptions to match your plan.