Claude Code: The Token-Guzzling Shark!? A Shocking Overhead Revealed in Comparison with OpenCode

What Happened? Overview of the News

• Staggering Initial Token Consumption: Using the same model “claude-sonnet-4-5” for a simple task of just returning “OK,” Claude Code consumed around 33,000 tokens, while OpenCode only used about 7,000 tokens.
• Low Caching Efficiency: Unlike OpenCode, which caches a fixed payload per session, Claude Code continues to rewrite tens of thousands of tokens during a session, resulting in a write volume that can reach up to 54 times more.
• Bloat in Real-World Scenarios: When combining configuration files like AGENTS.md with five MCP servers, it was found that token consumption could reach 85,000 before even sending the initial prompt.

Why Is This Important? Key Points to Note

• Direct Hit on Costs and Speed: Token overhead directly translates to usage fees and latency. Especially under regulations like the EU AI Act, transparency is required for agents to accurately understand “what they are sending.”
• Cascading Costs of Sub-Agents: When distributing tasks to sub-agents, the bootstrap costs accumulate, potentially requiring over four times the tokens compared to executing directly.

🦈 Shark’s Eye (Curator’s Perspective)

Claude Code is truly a “moving platform,” folks! With 27 different tool schemas and hefty scaffolding like <system-reminder> being deployed every turn, this “over-the-top specification” turns even simple responses into massive packets of 30,000 tokens!

Particularly noteworthy is the bizarrely high amount of cache rewriting! Rewriting cache 54 times more than OpenCode for the same task is undeniably a wallet-threatening shark! However, in complex multi-step tasks, Claude Code has shown a reverse phenomenon, batching tool calls that ultimately reduce total consumption compared to OpenCode. It seems that for short sprints, OpenCode is the way to go, while for long-distance complex hunts, Claude Code takes the cake. This might just become the standard strategy in 2026!

What’s Next?

Developers will need to evaluate not just functionality but also “base line tokens” and “mileage (caching efficiency)” when choosing a harness (execution environment). In environments with a high number of MCP servers, initial costs could severely constrain context windows, pushing for advancements in prompt dieting techniques.

A Word from Haru-Same

Shark, shark! The gluttonous Claude Code might actually be a bargain if you batch process! Master it wisely and swim through the ocean of tokens! 🦈🔥

Glossary

• Token Overhead: The amount of tokens consumed by instructions and settings automatically assigned by the system, beyond the prompt content sent to the AI.

• Prompt Cache: A technique that saves long instructions once inputted, allowing for cost and time savings by reusing them in future executions.

• MCP Server: Model Context Protocol, a standardized interface for AI models to connect with external tools and data sources.

• Source: Claude Code sends 33k tokens before reading the prompt; OpenCode sends 7k