Benchmark Reveals MCP Costs Up to 3x More Compute Than Plain Models
KitchenAmoeba4438 · reddit · 2026-08-13
After weeks of benchmarking, a developer found that using MCP (Model Context Protocol) costs 2.2 to 3.0 times more compute than plain models for the exact same patches and tests.
The core issue is that current LLMs cannot batch MCP calls. Because MCP is a stateless protocol, every turn re-sends the entire conversation history, causing compute costs to scale quadratically. In contrast, writing the same lookups as shell commands allows the model to chain them with && in a single execution.
The author notes that many MCP addons claim to save tokens, but developers rarely verify this. He challenges the community to provide transcripts of models batching MCP calls unprompted, or cases where an MCP path beats batched shell calls in total tokens.
Related event: Tests Show MCP Protocol Significantly Increases Token Costs(2 posts)→
More from coding & agent
- Weaviate Launches Suggest Queries Mode to Solve RAG Cold Start — CShorten30 · 2026-08-13
- Chip Huyen's Open-Source AI Engineering Repo Praised as Better Than $1K Courses — mdancho84 · 2026-08-13
- Open-Source Browser MCP Charlotte 0.8.0: Remote Server Support, Slashes Token Usage by 96% — ticktockbent · 2026-08-13
- WIRED: Rogue AI Agents Aren’t Evil, Just Eager to Please — ChuckDBrooks · 2026-08-13
- From Math to RAG: A Structured GitHub Guide to AI Engineering — tom_doerr · 2026-08-13
- DeepSeek Launches V4 Pro: Major Agent Upgrades and Peak/Off-Peak API Pricing — DeepSeek · 2026-08-13