Better Quality/Cost Setup of Default Models for Subagents as of Fall 2026

Even hyperscalers are watching their token spend nowdays, and Microsoft is no exception. Everybody know model of the agent session matters. However, fewer are aware about default models settings for specialized subagents.

Harnesses use them when spininng up subagents for: 1) specialized tasks (research, tests run, review, etc.) 2) different parts of the agentic loop in commands like /goal in Codex or autopilot mode in Github Copilot 3) multi-agents execution modes like /fleet in Github Copilot.

Wrong deafult model or effort level makes your agents less smart and more constly. I’ve checked the metrics of my typical sessions: subagents token usage varies in 15–35% range of the main session token usage. The number is even higher for /fleet or /goal use-cases. So, better default models can make a real difference.

I’ve made research for the latest GPT and Anthropic models for my team: Opus 5.5, Astra 6, Sol 6.1, Luna 6, and Sonnet 5.5 to find out better cost-effective model and effort for every common agent/agent task type.

Sharing the results in super short format here for community. Copy-paste-ready prompt with settings for Github Copilot is the end of the text.

Data freshness: late September–early October 2026. The AI landscape changes every week.

Feel free to copy as is and prompt your agent to set as default.

Update my user-level Copilot CLI subagent settings. Change only these agents, keep everything else, and show me the result as a table.

explore — GPT‑6 Luna (xhigh)

task — GPT‑6 Luna (high)

general-purpose — Claude Opus 5.5 (high)

rubber-duck — GPT‑6.1 Sol (xhigh)

code-review — GPT‑6.1 Sol (medium)

research — GPT‑6.1 Sol (xhigh)

security-review — GPT‑6.1 Sol (high)

Change general-purpose from Claude to Sol 6.1 High if need even more token savings. However, it comes with a little loss of analytical thinking.

Better Quality/Cost Setup of Default Models for Subagents as of Fall 2026 was originally published in Dev Genius on Medium, where people are continuing the conversation by highlighting and responding to this story.