AI Daily — July 13, 2026

Tools & Open Source

Claude Code Ships 33k Tokens of Scaffolding Before Your Prompt. OpenCode Ships 7k — Systima, a UK consultancy that sells LLM cost optimization, proxied two coding harnesses on the same machine and model and captured 150 request/response pairs. On a 22-character prompt, Claude Code sent roughly 33,000 tokens of scaffolding before the prompt arrived. OpenCode sent about 7,000, with tool schemas dominating both. Cache behaviour is the sharper finding, since Systima reports OpenCode's prefix stayed byte-identical while Claude Code re-wrote its full prefix mid-session, producing 5.9x to 54x the cache-write volume on a matched task. Configuration outweighs the harness floor either way, with a 72KB instruction file adding about 20,000 tokens per request and a two-subagent fan-out taking one task from 121,000 to 513,000 tokens. One result cut the other way, as Claude Code's batching made its multi-step total slightly lower than OpenCode's serialized nine-turn loop. Systima ↗

My takeaway: There were certain limitations to this experiment, and it needs more test cases to back it up. However, the method still holds. Log what you actually send to model. Measure how tokens are consumed across your own workloads. Then compare harnesses and configurations to find where the efficiency can be made.

Summaries are AI-generated and may contain errors — always verify against the linked original. Each story links to its source, which holds the copyright. Outlet names are shown for attribution only and do not imply any endorsement or affiliation.

Disclaimer: The views expressed in My Takeaway are my own personal opinions and general observations on industry trends. They are not intended to criticize, disparage, or make factual claims about any specific company, product, or platform. Any platform names mentioned are referenced solely for illustrative and informational purposes.