I pay for GitHub Copilot Business. Nineteen dollars a month. I thought that was the deal — I pay, I get Copilot.

Then one day I got a message saying I'd run out of something called "premium requests."

Wait, what?

If you're an engineering manager or CTO deploying Copilot across a team, you need to understand what you're actually paying for, because the bill at the end of the month might not be what you expected. I made a video walking through all of this, and this post covers the same ground for those who prefer to read.

The Era of Free AI Is Ending

For the last couple of years, AI has felt basically free. You sign up, you get access to incredible models, nobody asks hard questions about the bill.

That era is ending. The models are getting more powerful and more expensive to run, and the companies building them are passing those costs through. Here's a term you're going to start hearing a lot more: inference cost — the cost of running an AI model to generate a response. More powerful models cost more per request.

Inference cost management — understanding what your team is spending on AI compute and whether that spend is proportional to the value — is becoming a core competency for engineering leaders. And it starts with understanding how your AI tools actually bill you.

The Plan Landscape

GitHub offers six Copilot plans:

If you're managing a team, you're looking at Business or Enterprise. The big difference: Business gives you 300 premium requests per user per month. Enterprise gives you over three times that at 1,000.

"But I Already Pay for Claude Code"

This is the question every developer on your team is already asking. Why do they need a Copilot Business seat when they already have Claude Code or ChatGPT on their personal card?

Fair question. Those are great tools. But that's their account, their credit card, and their data governance decisions.

When a developer pastes company code into their personal AI account, that's a compliance conversation waiting to happen. And as a manager, you have zero visibility into what's being used, how much, or which models are being hit.

Copilot Business and Enterprise give you centralized billing, usage dashboards, audit logs, content exclusion policies, and control over which AI models your team can access. The value proposition isn't that the AI is better. It's that you get a managed AI platform with guardrails. The premium request system is how that management works.

The Two-Tier Model System

On any paid Copilot plan, the base models — currently GPT-5 mini, GPT-4.1, and GPT-4o — are included. Unlimited. No premium requests consumed.

But the moment you switch to an advanced model like Claude Sonnet, Claude Opus, Gemini Pro, or GPT-5.1, each interaction eats into your monthly premium request allowance. And not all models cost the same.

Model Multipliers: Where It Gets Interesting

There are over 20 models available in Copilot now. Here are the five tiers that matter for planning:

A model with a 3x multiplier costs three premium requests per interaction. A model at 0.33x costs one-third. (The full multiplier table has all 20+ models if you want the complete list.)

Here's what that means for a Business plan user with 300 premium requests per month:

• Gemini Flash (0.33x): 900 interactions — roughly 45 per working day
• Claude Sonnet (1x): 300 interactions — about 15 per day
• Claude Opus (3x): 100 interactions — 5 per day
• Opus fast mode (30x): 10 interactions — for the entire month

Someone could burn through an entire month's allowance before lunch on day one if they don't understand what fast mode costs.

What Happens When You Hit Zero

By default, Copilot falls back to the included base models for the rest of the month. The developer doesn't get cut off entirely, they just lose access to the advanced models.

Alternatively, organizations can enable paid overage at $0.04 per request. That doesn't sound like much, but if 50 developers each go 200 requests over, that's $400 in unexpected charges.

One cost management tool: Copilot's auto model selection gives you a 10% discount on multipliers when you let Copilot choose the model instead of picking one manually. Sonnet goes from 1x to 0.9x. Across a big team over a year, that adds up.

Policy Settings You Need to Configure

Before you roll Copilot out to your team, you need to make three decisions:

1. Premium request paid usage toggle. This controls whether developers can go over their allowance and trigger per-request charges. Decide before someone discovers Opus fast mode.

2. Budget caps. You can set a maximum spend on premium request overages. Without it, there's technically no ceiling on what a large team could spend.

3. Model access controls. You can control which AI models your developers are allowed to use. If you don't want anyone touching the 30x fast mode model, turn it off at the org level. This is probably the most underused policy setting in Copilot right now.

Real Cost: Business vs. Enterprise

At base cost, Business looks like the obvious choice. But as developers start hitting their 300 request cap and triggering overages, the gap shrinks. A thousand-dollar monthly gap becomes six hundred, then four hundred. And that's before you factor in the admin overhead of managing overage policies and developers getting cut off mid-month.

One More Thing: Spark and Coding Agent

Two Copilot features now track premium requests in separate billing buckets:

• Spark uses 4 premium requests per prompt, tracked in its own SKU
• The coding agent uses 1 premium request per session (multiplied by model rate), also tracked separately

This is actually good news for budgeting — you can see exactly how much your team is spending on agentic work versus regular Copilot usage.

The Question Nobody's Asking

If your team is burning through premium requests, the question isn't "do we need a bigger plan?"

The real question is: why are we throwing expensive models at tasks that cheaper models could handle just fine?

When a developer defaults to Claude Opus for every question, they're not making a cost decision. They're making a requirements decision — they just don't realize it. They're saying "every task I do requires the most powerful model." And that's almost never true.

Most code completions, most chat queries, most everyday development work can be handled by Sonnet or even the base models. Picking the right model for the task is really just understanding what you're asking for and why.

Model routing is requirements discipline wearing an AI hat.

Summary

1. Audit your policy settings. Is premium request paid usage enabled? Do you have a budget cap? Do you know which models your team can access?

2. Educate your team on multipliers. Most developers have no idea that switching from Sonnet to Opus triples their consumption. A five-minute conversation saves you hundreds of dollars a month.

3. Default to auto model selection. Let Copilot pick the model. You get the 10% discount, and for most tasks, Copilot picks well.

4. Track usage before upgrading plans. Before you go from Business to Enterprise, look at your actual usage data. If most of your team isn't hitting their cap, you don't need the upgrade. If they are, figure out why before you throw money at a bigger plan.

-Ben