Claude Code Injects 33k Tokens Before User Prompt, Sparking Context Opacity Debate
frequentcrazy33 · reddit · 2026-08-01
A developer compared the actual request overhead of Claude Code and the open-source tool OpenCode. Results show that in the first turn, Claude Code injects roughly 33k tokens of system instructions, tool definitions, and context before the user prompt, with 24k just for tool schemas. OpenCode's baseline is closer to 7k.
The author notes that while Claude Code's initial context is larger, it supports batched tool calls, potentially making fewer requests and costing less overall in multi-step tasks. However, this opaque context assembly prompted the author to rethink tool selection: choosing between convenient hosted agents that obscure context details, or self-managed open-source tools (like OpenCode, Aider, Cline) that offer transparency at the cost of manual setup.
More from coding & agent
- Deep Dive into AI Coding: Three Implementation Tiers of Spec-Driven Development — mattpocockuk · 2026-08-01
- Expert Rethinks AI Coding: Specs Should Be Ephemeral, Not Maintained — mattpocockuk · 2026-08-01
- Hidden Claude Code Setting Enables Infinite Retries for Long Tasks — mobileraj · 2026-08-01
- Ethiack Launches Agentic AI Pentesting, Claims 30x Speed Over Manual Tests — rez0__ · 2026-08-01
- Hugging Face Team Builds Agent-Generated Wiki for Reinforcement Learning — lvwerra · 2026-08-01
- Building a Stylized Grass Painting Tool with Codex and Three.js — anselm · 2026-08-01