How much context does Claude Code load before you type?
In the desktop app sessions we observed, the first request carried roughly 64,000 to 75,000 tokens before the prompt. That setup had several plugins, skills and MCP servers installed. A setup with fewer add-ons should start smaller. We haven’t measured the CLI or VS Code extension, so this post doesn’t compare them.
The example below rounds the starting context to 65,000 tokens. In our observations, most of the first request was a cache read: part of the starting context matched a recent session and was already cached. The rest was written to the cache. The example uses 40,000 tokens read and 25,000 written.
To see what fills your own starting context, run /context at the start of a session. It lists the system prompt, tools, MCP tools, memory files and skills separately.
What does starting context cost on a small task?
Suppose you ask Claude Code to fix one failing test in a small project. It reads a few files, runs the tests twice, edits one file, and finishes in 30 model requests. The task itself adds about 35,000 tokens of context by the end.
The table prices those requests at Claude Opus 5.5 API rates per million tokens, with five-minute cache writes. Check Anthropic’s pricing page for current rates and for one-hour cache writes.
| What was processed | Token type | Tokens | Rate per 1M | API cost |
|---|---|---|---|---|
| Starting context, first request | Cache write | 25,000 | $5.00 | $0.125 |
| Starting context, first request | Cache read | 40,000 | $0.20 | $0.008 |
| Starting context, re-read by 29 later requests | Cache read | 1,885,000 | $0.20 | $0.377 |
| Files, test output and edits | Cache write | 35,000 | $5.00 | $0.175 |
| Files, test output and edits, re-read | Cache read | 525,000 | $0.20 | $0.105 |
| New messages | Fresh input | 3,000 | $4.00 | $0.012 |
| Replies, tool calls and patches | Output | 12,000 | $20.00 | $0.240 |
The starting context accounts for $0.51 of the $1.04, about 49%. The work you asked for, including reading files, running tests and writing the fix, accounts for the other $0.53.
Why does starting context cost more than once?
Each model request sends the whole conversation, starting context included. Prompt caching makes the repeats cheap: in this example, re-reading 65,000 cached tokens costs 1.3 cents per request. But it happens on every request. Over 29 more requests, re-reads cost $0.38, nearly three times what the first request paid to load the context.
The cache also expires. If you step away and come back after it expires, the next request writes the context to the cache again at the higher rate. Anthropic’s prompt caching guide.
How much does trimming starting context save?
At these rates, each 10,000 tokens of starting context costs about $0.05 to write once and $0.002 for each later request. Over a 30-request task, that’s about 11 cents per 10,000 tokens.
Across a team, those cents add up. If a team runs 200 tasks like this a day, the starting context alone comes to about $102 a day at API rates, or about $2,100 over 21 working days. Trimming 20,000 tokens would save about 22 cents per task, or about $900 a month at that volume.
- MCP servers. Disable servers a project doesn’t use. Each one adds its tool descriptions to every session.
- CLAUDE.md. Keep the instructions Claude needs on every task. Move reference material into files it can read when needed.
- Plugins and skills. Remove the ones you don’t use.
- Sessions. Start a new session for an unrelated task instead of carrying old context forward.
Measure before and after with the same first prompt. Keep a change only if your tasks still succeed. A missing instruction that causes one extra retry can cost more than it saved. Claude Code cost management guide.
How do I check my own numbers?
Claude Code saves each session as a JSON Lines file under ~/.claude/projects. Find the first assistant message in a session and add its cache_creation_input_tokens, cache_read_input_tokens and input_tokens. That total is your first request, including your first prompt. To read the rest of the session, follow how to read a Claude Code session log.
For a team, use Claude Code’s OpenTelemetry data instead of local files. Track Claude Code cost per developer.
What does this mean on a Pro or Max plan?
A subscription doesn’t charge per token, so the dollar figures estimate what the same requests would cost at API rates. Starting context still counts toward your context window and usage limits. A larger starting context leaves less room for the task and makes Claude compact the conversation sooner. Check /usage for your plan limits.