If you’re using GitHub Copilot inside Visual Studio or Visual Studio Code, you’ve probably experimented with different AI models.
The big question:
Do you optimize for raw intelligence… or workflow speed?
After using Copilot daily for C# and .NET development, here’s my take.
Anthropic Models Are a Safe Bet
In general, models from Anthropic (Claude variants like Sonnet and Haiku) consistently generate:
- Cleaner code
- Fewer obvious bugs
- Better structural decisions
- Less manual correction
It really feels like software engineering is a priority in their training.
If you’re on Copilot Premium, Agent Mode in Visual Studio makes this even more powerful — you can scaffold solutions, create projects, install NuGet packages, and even run structured code reviews.
The Real Issue with Free Models: Speed
Free models can handle most tasks.
But they’re slow.
In my own testing inside Visual Studio 2026, even simple scaffolding tasks were roughly 2–3x slower. That delay kills momentum — and momentum is everything when coding.
AI only helps if it keeps you in flow.
If I Had to Pick One Model
Right now, my default model is:
Claude Haiku 4.5
Why?
- It handles about 80% of my daily coding tasks
- It’s significantly faster than free models
- It costs about one-third of Claude Sonnet 4.5
- It preserves my premium request quota
When I use higher-end premium models for 6 hours a day in .NET 10, I can burn through my entire monthly quota in 10–15 days.
Haiku dramatically reduces that.
Speed Beats Power (Most of the Time)
When rewriting a large codebase, AI should be able to:
- Create a solution
- Add multiple projects
- Wire up references
- Install packages
Free models often struggle just to scaffold a basic shell.
Premium “top-tier” models handle it instantly — but using them for routine tasks can be overkill and expensive.
Most development work isn’t advanced algorithmic reasoning.
It’s execution.
And in those cases, speed matters more than raw intelligence.
My Strategy
- Default: Claude Haiku 4.5
- Escalate: Switch to stronger models when stuck
- Use free models: Only when speed isn’t critical
For me, Haiku 4.5 hits the sweet spot between performance, quality, and cost.
What’s your default Copilot model right now — are you optimizing for power, speed, or budget?
Top comments (0)