New GitHub Copilot AI Licenses Dropped

It’s 5:44 PM PT exactly as I write this. I was watching the clock today—like a hawk—to see how the new GitHub Copilot AI credits were going to look, and boy, was I surprised. Here are my first impressions.

Credits Go by Fast

The first thing I did at 5:05 PM PT was ask Claude Sonnet 4.6 a troubleshooting question I was running into with a makefile. Nothing major, the usual:

“I’m just being lazy right now, please fix it virtual tech bro!”

Well, it turns out my virtual tech bro is trying to join the tres commas club, and in true Russ Hanneman fashion, every chain-of-thought, every tool call, and every file read/write was eating into those sweet, sweet credits.

When Claude finally returned an answer to a very basic problem, I had already burned through 61 credits of my 7,000 monthly budget. That’s a ton of usage for not a lot of gain.

New VS Code Changes

Adjustable Thinking Levels

The one nice change is the ability to adjust the level of thinking across models. This used to be available in the early days, but they removed it in favor of hardcoding specific thinking modes. Claude Sonnet 4.6 (High), my favorite workhorse in the stable, is no longer locked to a single mode:

Now you can switch to a thinking mode that works best for your situation. I’m already dropping mine down to Medium in hopes of reducing AI credit consumption:

Changes to the Model Lineup

The freebie models are gone. The days of having GPT-4o mini available for zero tokens are over. Now, every single available model consumes credits:

Copilot Max Subscription

You can now upgrade to the Max subscription, which offers 20,000 AI credits a month. Considering how quickly credits vanish for simple tasks, it’s worth considering. Like the sane, rational human being that I am, I immediately upgraded the second I saw the burn rate.

I cannot wait to try an MCP server…

The Easy Days Might Be Over

This might be the end of the high-flying days I’ve been experiencing the last few months with GitHub Copilot Plus. Fortunately, I crammed as many features as I could into my projects before the credit system started. I’ll miss getting features done without burning through requests, but hey, maybe slowing things down a bit is needed?

Forget that noise!

Local Models to the Rescue

I have really been digging qwen3.6-27B with thinking enabled on mlx_vlm. It has been the powerhouse for my MLOps project ferme, which is currently generating Alpaca pairs to fine-tune a gemma4 and qwen2.5-coder model using QLorA.

My M1 Max 64GB is a tight fit, and its TPS is not the greatest, but once my synthetic seeding is done and I have all these Alpaca pairs, it might be time to repurpose qwen3.6 for coding tasks and gemma4 for simple generation and documentation tasks. Sounds like it’s time to get serious about integrating Qwen Code directly into my VS Code workflow!