Learn the models.
Spend wisely.

Use a strong model to plan. Use a lean model to build. Know when to switch.

Find your model Explore the news

Give each model the right job.

A starting strategy, not a rule: spend more where judgment matters and less where the steps are clear.

01

Plan

Use GPT-6.1 Sol or another stronger model to clarify goals, compare approaches, and write testable steps.

Deliverable: a bounded plan + acceptance checks.
02

Build

Try GPT-6 Luna for focused edits, documentation, and implementation against that plan.

Deliverable: small changes + passing checks.
03

Review

Use a stronger model for risky decisions or unresolved failures. Verify results with tests and human judgment.

Escalate when the same failure repeats.
What should I hand to the cheaper model?

Give it the goal, relevant files, the agreed approach, constraints, and commands or checks that prove success. Work in small steps. Ask it to stop and report a blocker instead of guessing.

“Implement step 2 of this plan. Keep the public API unchanged. Run the listed checks. If the assumptions fail, explain what you found before changing the approach.”

Switch models in the Copilot model picker. Preserve the plan and relevant context when starting a new conversation; switching alone does not guarantee an effective handoff.

Start with the task.

Choose a situation to see a practical starting point.

These are editorial recommendations, not model benchmarks. Judge the actual result, not the model’s confidence.

Same tokens. Different price.

20,000 uncached input tokens + 4,000 output tokens, at standard rates.

Prices checked October 1, 2026

Linear scale, $0 to $0.400. Gemini 3.8 Flash uses its promotional rate through December 31, 2026. This compares prices for equal tokens, not quality or cost per successful task. Input tokens are what you send; output tokens are what the model generates, including billed reasoning.

GitHub’s official rates · AI-credit billing: 1 credit = $0.01. Subscription fees, included allowances, caching, and long-context tiers are excluded here. Availability depends on your plan, client, and organization policy. Legacy annual request-based plans differ.

Make the tradeoff visible.

Adjust the work. See what model switching could save.

Adjust the assumptions.

Overhead is your assumption, not a fixed “High” multiplier. Higher thinking effort can consume more tokens. Extra attempts repeat both input and output.

Estimated usage cost

Sol for every step$0.480
Sol plan + Luna build$0.100
79%less estimated usage cost

Illustrative estimate, not a benchmark. Both routes include the same planning + review steps. Every step uses the token counts you enter; real steps vary.

See the calculation

Cost = (input tokens × input rate + output tokens × output rate) ÷ 1,000,000. No cache reads or writes. Standard tier only; per-step input is capped at 272K. Dollars show usage value, not necessarily an extra charge on your bill.

How thinking effort affects usage

“High” is a setting. Success is the goal.

Start with Luna on a bounded task. If your client offers thinking effort, try a higher setting when deeper reasoning is useful. There is no published universal conversion from “High” to a fixed cost.

Check the result. If it repeats a failed approach, misses constraints, or needs architectural judgment, pause and switch. A cheap attempt that never finishes is not an efficient workflow.

Read more: GitHub’s model comparison · Full pricing reference

Keep up. Keep building.

See all news →

Workflow Patterns

Real AI Workflows

Capture practical ways your team is using agents, skills, and AI-assisted tooling inside VS Code.

RAG-style Log Troubleshooting Agent Using a Reference Folder Agent + Skills

Build a dedicated Copilot agent that uses a reference folder of documentation you choose and can alter, incident notes, and snippets of common issues to troubleshoot new logs faster with retrieval-style context.

How To Set It Up

  1. Create a reference/ folder with the exact reference documents you want it to use, known error signatures, and resolution notes.
  2. Add a custom .agent.md focused on troubleshooting and root-cause analysis.
  3. Create one or more SKILL.md files for parsing logs, matching patterns, and proposing fixes.
  4. Attach those skills to the agent so it consistently follows your triage workflow.
  5. Prompt the agent with new logs and ask it to compare against the reference folder before recommending actions.

Why This Works Well

  • Keeps the agent focused, and if you add limits to the tools it can't attempt any rewrites.
  • Reduces repeated manual searching across old tickets and files.
  • You are able to prompt it to ensure the responses it gives meet your standards to reduce edits before passing information along, and to speed checking the results
  • You can ask for it to find links to the information in the documentation to support its answers

Copilot Build Assist

Copilot can help generate both the agent and skills: ask it to scaffold your .agent.md, draft SKILL.md instructions, and refine prompts for better troubleshooting output over time.

Tip: Use Copilot and explain what you want the agent to focus on. After it writes the agent for you, ask Copilot if there are any useful skills it sees that might help the agent function better. It is amazing what it can come up with to help itself improve.

Improving Reference Materials Efficiency for the Agent Optimization

You can take the reference folder you created, and prompt the agent to go through the files and restructure them for its efficiency. Create new documents to document types that are more efficient, and rename or reorganize to improve retrieval, and use less tokens. This is a great way to optimize the agent's performance without changing the underlying code of the agent or skills, just by improving the materials it uses to reference.


Tip: Make sure you tell the agent not to 'rewrite' the information, but to only restructure it. You don't want to lose any of the important information, just make it more efficient for the agent to reference.

Core Concepts

Agents and Skills at a Glance

These are the reusable building blocks behind more advanced AI workflows in VS Code.

Agents

  • Autonomous coding assistants that work directly on GitHub issues
  • Defined via .agent.md files in your repository
  • Open pull requests with proposed changes for team review
  • Customizable with specialized instructions and attached skills

Skills

  • Reusable instruction sets written in SKILL.md format
  • Stored in .github/skills/ (project) or ~/.copilot/skills/ (personal)
  • Attach to agents to give them specialized capabilities
  • Shareable across repositories and team members

In Practice

Using AI in VS Code

A strong setup balances quick chat loops, deeper agent execution, and repeatable automation.

Interactive Help

Use chat for quick explanations, code edits, documentation lookup, and small in-context fixes without leaving the editor.

Agentic Execution

Use agent mode when a task spans multiple files, needs terminal commands, or requires validation rather than one-shot suggestions.

Repeatable Automation

Keep recurring content in a standalone data file so scheduled jobs can update the feed without rewriting page layout or UI logic.

This site now uses news-feed.js for that pattern.

Learning Path

Getting Started

Follow these steps to go from zero to productive with AI in VS Code.

1

Understand the Basics

Learn what agent skills are and how they fit into the Copilot ecosystem.

About Agent Skills
2

Try Agent Mode

Get hands-on with agent mode through a guided Microsoft Learn tutorial.

Building Apps with Agent Mode
3

Create Your First Custom Agent

Define a specialized agent with its own identity and instructions.

Creating Custom Agents
4

Build Reusable Skills

Create SKILL.md files that give agents specialized capabilities your whole team can use.

Creating Agent Skills
5

Master Best Practices

Learn proven patterns from 2,500+ repositories and official GitHub guidance.

Best Practices How to Write a Great agents.md

Cheat Sheet

Key Files & Locations

.agent.md Custom agent definition file — identity, instructions, and skill attachments
SKILL.md Skill instruction file — reusable capability definitions
copilot-instructions.md Repo-wide Copilot instructions — applies to all agents
.github/skills/ Project-level skills directory — shared with the team via repo
~/.copilot/skills/ Personal skills directory — your private skill library
news-feed.js Shared AI news data source — update this file to refresh both the homepage snapshot and the full tracker

Model Guide

Go Deeper in the Model Guide

The model table is now a dedicated reference with current Copilot multipliers, task fit, and recommended model flows for planning, implementation, review, and quick edits.

Cost-aware model table

Compare selected models, token rates, availability notes, and suggested starting points.

Task recommendations

Pick a practical flow for planning, implementation, debugging, refactoring, docs, visual work, and low-cost loops.

Official-source notes

Includes GitHub's current model, billing, and auto-selection caveats so the guide is easier to maintain.

Open the full guide: GitHub Copilot Model Guide

Deep Dive

AI in VS Code Library

Curated docs, tutorials, repositories, and reference material for agents, skills, extensions, and related AI workflows.