---
title: 'In the News: November 1, 2026 (Extra 3)'
description: 'SineFrame reports Claude Haiku 5.5 scored 82.0% on 21 MCP servers through OpenCode and 65.4% through Claude Code, driven by built-in tool overlap.'
canonical_url: 'https://darkfactory.dev/news/2026-11-01-extra-3'
markdown_url: 'https://darkfactory.dev/news/2026-11-01-extra-3.md'
collection: news
date_published: '2026-11-01T23:10:00-05:00'
date_modified: '2026-11-01T23:10:00-05:00'
---

# In the News: November 1, 2026 (Extra 3)


One story this cycle: a vendor benchmark reporting that the coding agent wrapped around a model changes whether an MCP server gets called at all.

## 1. The same model scored 17 points apart on 21 MCP servers, depending on the coding agent

**[Same model, different agent: 17 points apart on 21 MCP servers](https://m3.sineframe.com/blog/same-model-different-agent)** · Utkarsh Pandey, co-founder · SineFrame, November 1, 2026

SineFrame, which builds the open-source MCP testing tool M3, reports that Claude Haiku 5.5 answered 82.0% of tasks correctly from the server through OpenCode and 65.4% through Claude Code. The runs covered 21 MCP servers, about 20 tasks each, three trials per task, 3,717 scored trials between October 29 and November 1. Codex on GPT-6 Luna scored 56.5%, though that column changes the model as well as the agent. The gap sits on ten servers that overlap with an agent's built-in tools, such as web search and files: Claude Code scored 34.5% there against 66.3% for OpenCode. On the eleven servers with no built-in overlap, the three agents landed within 4.1 points of each other.

A trial counts as a success only if the answer is correct and the key value appears in a server response, so a right answer found through the agent's own web search is scored a miss. SineFrame says Claude Code answered 378 overlapping-server trials through its own tools, and 342 of those answers contained the expected value. It reports that Codex skipped the server on 44.5% of trials where it should have called it, Claude Code on 36.5% and OpenCode on 16.1%. In 1,239 trials OpenCode never called a built-in web search tool. These are the company's own figures; we have not rerun them, and the author sells the tool the benchmark was built to exercise.

**Why it matters:** A server can work and still never be called, because it competes with the tools the agent brings along. If you ship or rely on an MCP server, measure how often the agent chooses it, not only whether the final answer is right, and test through the agent your users actually run.