Claude 3.7 Sonnet, Gemini 2.5 Pro, GPT-4… developer choice is key to GitHub Copilot, and that’s especially true when it comes to picking your frontier model of choice. 

But with so many frontier generative AI models now available to use with GitHub Copilot (and more coming seemingly every day), how do you pick the right one for the job—especially with the growing capabilities of Copilot Chat, edit, ask, and agent modes?

In a recent video, I worked with GitHub’s Developer Advocate Kedasha Kerr (aka @ladykerr) to answer this exact question. Our goal? To build the same travel‑reservation app three different ways with Copilot ask, edit, and agent modes while swapping between Copilot’s growing roster of foundation models to compare each AI model in real-world development workflows. 

We set out to build a very simple travel‑reservation web app (think “browse hotel rooms, pick dates, book a room”). To keep the demo snappy, we chose a lightweight stack:

• Backend: Flask REST API
• Frontend: Vue.js, styled with Tailwind
• Data: a local  data.json file instead of a real database

That gave us just enough surface area to compare models while scaffolding the app, wiring up endpoints, and adding tests, docs, and security tweaks along the way . 

Here are a few key takeaways from our video (which you should watch). 

But first, let’s talk about Copilot’s three modes

GitHub Copilot gives you three distinct “modes:” ask, edit, and agent mode.  Ask is there to answer questions, edit is a precise code‑rewriting scalpel, and agent mode can drive an entire task from your prompt to the finished commit. Think of it this way: Ask answers, edit assists, agent executes. 

Tip 1: No matter what model you use, context matters more than you think

The model you use is far from the only variable, and the context you offer your model of choice is often one of the most important elements. 

That means the way you shape your prompt—and the context you provide Copilot with your prompt and additional files—makes a big difference in output quality. By toggling between capabilities, such as Copilot agent or edit mode, and switching models mid-session, we explored how Copilot responds when fed just the right amount of detail—or when asked to think a few steps ahead.

Our demo underscores that different modes impact results, and thoughtful prompting can dramatically change a model’s behavior (especially in complex or ambiguous coding tasks). 

The takeaway: If you’re not shaping your prompts and context deliberately, you’re probably leaving performance on the table.

For a deeper dive into model choice, the guide “Which AI model should I use with GitHub Copilot?” offers a comprehensive breakdown.

Tip 2: Copilot agent mode is a powerful tool

Agent mode, which is still relatively new and evolving fast, allows Copilot to operate more autonomously by navigating files, making changes, and performing repository-wide tasks with minimal hand holding. 

This mode opens up new workflow possibilities (especially for repetitive or large-scale changes). But it also demands a different kind of trust and supervision. Seeing it in action helps demystify where it fits in your workflows.

Here are two ways we used agent mode in our demo: 

• One‑click project scaffolding: Kedasha highlighted the project README and simply told Copilot “implement this.” Agent mode (running Gemini 2.5 Pro) created the entire Flask and Vue repository with directories, boiler‑plate code, unit tests, and even seeded data. 
• End‑to‑end technical docs: I started using agent mode with Claude 3.5 and prompted: “Make documentation for this app … include workflow diagrams in Mermaid.” Copilot generated a polished README, API reference, and two Mermaid sequence/flow diagrams, then opened a preview so I could render the charts before committing .

Tip 3: Use custom instructions to set your ground rules

Another insight from the session is just how much mileage you can get from customizing Copilot’s behavior with custom instructions. 

If you don’t know, custom instructions let you lay down the rules before Copilot suggests anything (like how APIs need to be called, naming conventions, and style standards). 

Kedasha in particular underscored how custom instructions can tailor tone, code style, and task focus to fit your workflow—or your team’s. 

One example? Using custom instructions to give every model the same ground rules, so swaps between each model produced consistent, secure code without re‑explaining standards each time.

Whether you’re nudging Copilot to avoid over-explaining, stick to a certain stack, or adopt a consistent commenting voice, the customization options are more powerful than most people realize. If you haven’t personalized Copilot yet, try custom instructions (and check out our Docs on them to get started).

Tip 4: The balance between speed vs. output quality

No matter what model you use, there are always tradeoffs between responsiveness, completeness, and confidence. A larger model may not provide quick suggestions when you’re working through an edit, for instance—but a smaller model may not offer the best refactoring suggestions, even if it’s faster in practice. 

TL;DR: It’s not about chasing the “best” model—it’s about knowing when to switch, and why. Your default model might work 80% of the time—but having others on deck lets you handle edge cases faster and better.

Take this with you

This video demo isn’t a scripted feature demo. It’s two devs using Copilot the way you would—navigating unknowns, poking at what’s possible, and figuring out how to get better results by working smarter instead of harder.

If you’ve been sticking with the defaults or haven’t explored multi-model workflows, this is your invitation to take things further.

👉 Watch the full video to see how we put Copilot to work—and got more out of every mode, prompt, and model. 

Now—what will you build? Try GitHub Copilot to get started (we have a free tier that’s pretty great, too). 

Additional resources: 

• Explore the demo repository: Try forking our demo repository from the video to test our different models with GitHub Copilot. 
• Which AI model should I use with GitHub Copilot? A look into each model currently offered with GitHub Copilot from Cassidy Williams. 
• A guide to deciding what model to use in GitHub Copilot: A framework for figuring out which model to use when new models are appearing what feels like every day (or maybe every week). 

The post Real‑world video demo: Using different AI models in GitHub Copilot appeared first on The GitHub Blog.

If ever you've thought that your AI may have been listening in on your private conversations...

Well, it brings me no pleasure to bag on Pepsi but the people have spoken with their almighty dollars.

"The state postal service of Denmark, PostNord, has announced it will cease handling letters in Denmark at the end of 2025 "to focus on becoming the Danes’ favorite parcel delivery service." The cut in service is tied to a reported 90% decline in letter volumes since 2000.

Postnord has also announced that any postage stamps bought in 2024 or after could be refunded for a limited time in 2026, presumably marking the end of the country's stamp issues.

Denmark's Transport Minister, Thomas Danielsen, told Danes concerned about the future of their mail that "a free market" would ensure they could still send letters, just not through a national postal service...

An increase in use of digitial communication is largely cited as the reason for the decline in number of letters sent in the country. However, Denmark's Postal Act of 2024 also allowed private firms to enter the postal arena, and inland letters are no longer exempt from value-added tax, or VAT, meaning that a letter cost Danes 29 krone ($4.20) to send...

Exceptions were previously made for populations in remote areas and with vision impairment, but PostNord's statement does not suggest that will be the case again. The statement also does not directly say whether letters can be sent from other countries to Denmark; however, at this time we can only infer that international letters will not be accepted or distributed..."

VATICAN — In an aerial display that delighted millions, the Vatican conclave released a bald eagle firing an AR-15 into the air to signal that an American pope had been selected.