GitHub Copilot Evaluation Models Now Available for Individual Plans

TL;DR

• GitHub Copilot now serves evaluation models to individual (non-enterprise) users through auto model selection
• Evaluation models let you test new AI capabilities before they're fully released
• You can disable this in GitHub Copilot settings if you prefer stable models only

What Dropped

GitHub expanded access to evaluation models for Copilot individual plan users. These experimental models can now be automatically selected and served alongside your standard model choices, giving you early access to new AI capabilities that GitHub is testing.

The Dev Angle

Evaluation models are pre-release versions of AI models that GitHub wants developers to test in real workflows. They're not production-ready, but they let you kick the tires on upcoming features and improvements before official launch. If you're on an individual Copilot plan, you're now part of this testing pool by default.

The auto model selection feature intelligently routes requests to different models based on task complexity and your settings. With evaluation models enabled, Copilot may choose an experimental model if it thinks it's better suited for your current task. This is similar to how targeted model rules let you control which models handle specific scenarios.

You maintain full control: flip over to your GitHub Copilot settings and disable evaluation models if you want to stick with stable, battle-tested versions only. No surprises, no forced upgrades.

Should You Care?

If you're an individual Copilot user and you like living on the edge—testing new features, giving feedback, helping shape the product—this is a win. You get early access to improvements without paying enterprise rates.

If you need predictable, stable behavior and don't want experimental code suggestions, disable it immediately. One toggle in settings and you're back to standard models only. The choice is yours.

The real value here is transparency: GitHub is being explicit about when you're using evaluation models and giving you an easy off-ramp. That's better than silent A/B testing.

Source: GitHub Changelog