Where are we coming from? Just six days ago.

On April 22, I wrote that things were moving at a pace most people weren’t really noticing, at least in my circle. Anthropic, GitHub, and SAP all made their moves almost at the same time. The signal was quite clear: flat subscription models for AI were starting to break down. Not for users, but for the companies building these services.

Five days later, GitHub didn’t just confirm the direction, it put a date and a mechanism behind it. At that point, there’s not much left to interpret.

From PRUs to credits: the shift that confirms everything

Starting June 1, Premium Request Units disappear. They are replaced by GitHub AI Credits, consumed based on tokens — input, output, and cached — according to each model’s pricing.

The plan pricing doesn’t change. Copilot Pro remains at $10 per month, Business at $19 per user, Enterprise at $39. But the meaning changes: every dollar you pay now directly translates into usable credit. No buffer. No safety net.

That buffer used to be the fallback. When PRUs ran out, Copilot would continue working using a cheaper model, and the user wouldn’t notice. That safety net disappears on June 1. From that moment on, when credits are exhausted, what happens next depends entirely on the budget configured by the admin.

Why agents break flat subscriptions

GitHub explains it quite plainly: a multi-hour agent session and a quick chat prompt cost the same for the user. GitHub absorbed the difference.

With agent-based usage becoming the norm, that model is no longer sustainable.

This isn’t a complaint, it’s a technical reality. Agents don’t make a single model call. They typically make dozens, sometimes hundreds, chained together. They reason, iterate, validate, and retry. Every step consumes tokens. Charging the same for that as for a simple interaction doesn’t make economic sense in the long run.

SAP pointed to this in March. Anthropic acted on it in April. GitHub has now formalized it with a timeline. These are not isolated moves.

What changes for teams and organizations

Beyond the credit model itself, there are two operational shifts worth paying attention to.

The first is shared credit usage. Until now, unused credits from one user were simply lost. There was no way to redistribute them. With the new model, Business and Enterprise plans will include an organization-level shared pool. Low usage from some users can offset heavier usage from others. It looks like a small change on paper, but in practice it’s quite impactful.

The second is budget controls. Admins will be able to define budgets at the company level, cost center level, and individual user level. Once the included credits are consumed, the organization decides whether to allow additional spend at standard rates or cut off access. This is real AI cost governance, with a level of granularity that hasn’t been natively available before.

The part that concerns me the most

It’s not the pricing. Plans are not getting more expensive. For moderate usage, the change may barely be noticeable.

What stands out is the speed. In just a few months, we’ve moved from a kind of open-ended creative usage model to a metered consumption model, and we’ve done it with very little open discussion.

The public conversation is still focused on which model is better or which tool is more powerful. But the discussion about what it actually costs to use these tools in real scenarios — with agents, long-running sessions, autonomous pipelines — has barely started.

June 1 will make that visible for many people. What was a trend a few days ago is now a concrete date.

So this is the moment to pause and reassess: what tools you’re using, what they actually consume, and whether your budgets are designed for a chat-based world or an agent-driven one.

Because those are two very different worlds.

References

• GitHub Copilot is moving to usage-based billing — GitHub Blog
• Changes to GitHub Copilot Individual plans — GitHub Blog