Today I asked opus 4.6 - 2 questions and it already “felt” as if 60-70% of its context was used up. An interesting result Question 1: What is in the context window? Answer 1 [Opus 4.6]: Right now, the context window contains the following: System prompt — This is the main set of instructions that guides how I behave. It includes: ∙ Product information about Claude and Anthropic’s offerings ∙ Behavioral guidelines (tone, formatting, safety, copyright, evenhandedness, etc.) ∙ Web search instructions and copyright compliance rules ∙ Image search tool guidance ∙ Computer use instructions (access to a Linux container with bash, file creation/editing tools) ∙ Available skills (docx, pdf, pptx, xlsx, frontend-design, product-self-knowledge, skill-creator) ∙ Past chats tool instructions (for searching conversation history) ∙ Artifact creation guidelines (React, HTML, Markdown, etc.) ∙ Anthropic API-in-artifacts capability (for building AI-powered artifacts) ∙ Citation instructions for web search results ∙ Tool definitions for all available tools (web search, image search, file tools, places/maps, sports data, calendar/reminders, message compose, etc.) ∙ Network and filesystem configuration (network egress is currently disabled) Your message — Your single question: “what is in the context window right now” That’s it — there are no uploaded files, no prior conversation turns, no memory entries, and no images. It’s just the system instructions and your one message. Question 2: how heavy does that feel if you gave it a percentage? Answer 2 [Opus 4.6]: Honestly, it’s on the heavier side — I’d estimate around 60-70% of the context window is already occupied by the system prompt alone. There’s a lot in there: detailed behavioral guidelines, extensive copyright and search instructions, tool definitions for 20+ tools, skill descriptions, artifact rules, the API-in-artifacts documentation, past chats tool logic, and various safety/formatting policies. It’s quite a dense set of instructions. That leaves roughly 30-40% for our actual conversation, any files you might upload, tool call results (like web search responses), and my replies. For a typical back-and-forth chat that’s plenty, but if we were doing something like deep research with lots of search results or working with large files, it could start to feel tight over a long conversation. submitted by /u/dermflork
Originally posted by u/dermflork on r/ArtificialInteligence