Introduction
I have been working professionally as a software engineer since 1999 when I got my first job at Fujitsu Consulting. Since then I worked at various companies and held titles of technology consultant, software engineer, senior software engineer and principal software engineer as well as solutions architect. I have worked with AI for the past few years and have been doing Agentic development for about a year now. From my perspective manual coding is dead, AI can do coding much faster and better than I can.
Having said that I still think that building applications is not dead, we just build them differently. Today I would like to write my first article on Agentic development or developing using AI agents. For this article I am going to use the technology stack I am most familiar with: VS Code and GitHub Copilot. I have some experience with Open AI Codex, however I work with GitHub Copilot about six to eight hours a day so I have a lot more experience with it.
GitHub Copilot Plans
You can see the available plans here: https://github.com/features/copilot/plans
Which plan would I pick? I would recommend starting with Pro+ and if you follow my tips in this post you should be able to develop for at least four hours a day and not run out of credits. I use about $100 worth of credits every month working with GitHub copilot for 40 hours each week.
As you can see Pro+ gives you about $70 worth of AI credits every month and if you are very careful with how you use credits you can get away with $70 per month working 40 hour weeks on AI. It is also very easy to blow your monthy credit allowance even on the Max plan if you are using the most recent models especially those from Anthropic (Claude) and don't use the tips I am going to share below.
LLM Model Pricing
One of the most important things to consider when working with AI is what Models you are going to use. The newer the model the more expensive it tends to be, also the more thinking you are going to outsource to the LLM model the more expensive it is going to be.
So the consideration of choosing your tools is what LLM Models are available to you, as you can see if you pick Claude Code you are going to be locked into Anthropic Models only and same goes for OpenAI, while with GitHub Copilot you can chose what LLM providers you are going to use:
Choosing the right LLM Model for the task is a very important task of managing the cost of tokens. You can see the various costs for the models that GitHub Copilot uses here: https://docs.github.com/en/copilot/reference/copilot-billing/models-and-pricing
If you look at the table below you will see that costs for Anthropic models such as the latest Opus are obscene:
I have been excluding Claude Opus 4.8 and Opus 5 from my coding because the results were not that much better than Claude Sonnet 5 and price is much higher. Sonnet 5 is the most expensive LLM model I am currently using. And let me tell you, it consumes credits at a very fast rate compared to my most frequently used model: GPT 5.6 Luna
Luna is a very capable model if you are willing to do some thinking yourself instead of letting LLM do the thinking for you. Another model that I found to be quite capable and not very expensive was Google Gemini Flash 3.7 model, there is also 3.8 Flash which is now available but I have not used it much yet:
GPT-5.3 Codex used to be my go to model, but I find myself using it less and less.
So this is kind of the more high level view of what goes into the costs of Agentic development, in the next part of these series I am going to talk about more specific ways of getting the most value out of your AI tokens and note that while I am going to be using GitHub Copilot with VS Code as my examples these same principles apply to any Agentic development tool such as Claude Code and others.
Next post is going to show how I manage Context for cost optimization and how I approach this type of development with a specific Hive example.
I am a Hive Witness and would really appreciate your vote for me as a witness: https://vote.hive.uno/@seattlea
What is a model and why would you want to use a different one for different things?
:) Excellent question! You can think of LLM models as capability sets, different LLM models are good at different things. For example Claude Opus is great at thinking and solving complex problems, but it comes with a giant cost no matter what it does. So if I use Claude Opus in GitHub Copilot to connect to Azure MCP server and run simple log analysis it will cost me 250 tokens to do this task, but if I take GPT 5.6 Luna it will do the same task faster (because it will think a lot less) and for the cost of only 2 tokens...
P.S. these are real numbers based on my experience...
Okay thanks. Is there a general catch all if you don't really know what you want to do? Something middle of the road that covers a lot of things?
I don't think middle of the road is going to cut it :) I have a very detailed example post coming out tomorrow :)
Thank you!
You're welcome 🤗
Cheers to finishing your daily Ecency leaderboard quest, @seattlea! Your impressive two decades of software engineering experience and professional journey are truly inspiring.
Tipped with Ecency POINTS.