GitHub Copilot CLI Is a Hidden Gem
My son wanted to start using AI coding agents. He had watched enough YouTube to know roughly what the workflow looked like. He was ready to get started.
My professional experience almost sent him down the wrong path. I use Claude Code and Codex every day. That’s my frame of reference. My instinct was to just say “here, use what I use.” I caught myself.
When I actually sat down and looked at the options for someone starting out, GitHub Copilot was the obvious answer. $10 a month, access to frontier models, and a CLI that’s a full coding agent. I had been mentally filing Copilot under “IDE autocomplete perk” for years and never looked closely at what it had become.
My son, for his part, had seen videos about “Copilot” and coding agents and assumed they were talking about the same thing I was. They weren’t. There’s the autocomplete. There’s the chat. There’s the unfortunate Windows feature. And then there’s the CLI, which is awesome.
Here’s the thing. I run Claude Code and Codex every day at work and they’re excellent. But my familiarity with them is exactly why I almost gave bad advice. For personal use, the value prop isn’t close.
What You’re Actually Getting
A GitHub Copilot Pro subscription is $10 a month. Pro+ is $39.1 For that you get 300 and 1,500 premium requests per month respectively, plus access to a wide spread of frontier models: GPT-4.1, Claude Sonnet, Claude Haiku, Gemini Flash, and more.
Compare that to going direct. Claude’s API charges per token. Anthropic’s Max plan runs $100/month with hard caps. OpenAI is similar. For most developers who aren’t running agentic loops all day, the GitHub plan covers what you actually need at a fraction of the cost.
The free tier gives you 50 premium requests and 2,000 code completions. For someone just starting out, that’s real.
The Multiplier Model Is Honest
GitHub publishes their model multipliers.2 Claude Sonnet 4.6 and GPT-5.4: 1x. Claude Opus 4.6: 3x. Every request on your plan gets consumed at the stated rate for the model you chose. Gemini models come and go from the catalog but tend to sit at fractional cost, effectively multiplying your budget on lighter tasks.
No mystery. No hidden metering. You know exactly what you’re spending when you reach for a stronger model versus a faster one.
This means if you’re deliberate about model selection, your budget goes further. Sonnet for most things. Opus when you need it. You stretch 300 requests a long way if you’re not reaching for the heaviest model by reflex.
Per-Request Is the Key Unlock
GitHub charges per request, not per token. So if you construct a dense, well-structured prompt, you get a lot out of a single shot.
I’ve sent single requests that produce complete feature implementations, tests, and documentation. They cost one or five premium requests, depending on model. The same work going direct costs dozens of API calls and significantly more money.
This rewards the kind of careful prompting that good agentic work requires anyway. If you’re going to use an agent seriously, you need to front-load context, specify what you want precisely, and not iterate lazily. The pricing model just makes the incentive structure visible.
The CLI Is the Real Story
Here’s what I actually told my son about.
The GitHub Copilot CLI is a full coding agent.3 It understands your repository. You define custom agents via an AGENTS.md file at your repo root — each agent gets its own persona, tool access, and behavioral scope.4 Hand off work with /delegate and it creates a branch, implements the task in the background, and opens a pull request when it’s done. Model selection is a command: /model. Switch based on what the task actually needs.
The skills system extends this further. You can attach MCP servers, domain-specific knowledge, and custom tools to an agent. The agent I use to maintain this site knows the design language, the content schema, the component structure. I don’t explain any of that per request. It’s ambient context.
For a new developer, this is the right starting point. It lives in a place they already know (GitHub). Setup is trivial. The behavior is controlled and predictable. You don’t need to manage API keys, pick an agent framework, or develop prompting intuitions before writing any code.
Buried in the Bundle
I think the reason this gets missed is that Copilot is bundled with everything. It’s a GitHub feature. It’s the autocomplete in your IDE. It’s the thing your company already pays for. That framing makes it easy to underestimate. The CLI in particular is newer and gets almost no coverage relative to Claude Code or Cursor.
I missed it, and I work in this space. That’s not a great sign for how visible it is.
For value per dollar, especially at the $10 tier with the CLI as your primary interface, it holds up against anything I’ve used. Sometimes the best recommendation isn’t the most interesting one.
For my son, it was the right call.