Post
GitHub Copilot Just Started the Meter
On June 1, GitHub stops counting Copilot in 'premium requests' and starts counting it in retail tokens. The base prices didn't move, but the math underneath quietly did. Here's what flipped, and what it means for anyone who runs an agent.
So GitHub just announced something that sounds boring on the surface and is actually a meaningful pivot.
Starting June 1, 2026, Copilot stops measuring you in "premium request units" and starts measuring you in GitHub AI Credits — input tokens, output tokens, cached tokens, all priced at published API rates.
The base prices didn't move. Pro is still $10. Business is still $19. Enterprise is still $39. If you only read the headline, you'd think nothing happened.
But everything underneath changed.
What got swapped
The old model was a buffet with portion limits. You'd get N "premium requests" per month, and once you hit the quota, Copilot would silently downgrade you to a cheaper model so you'd keep getting answers — just less impressive ones.
The new model is a tab at a coffee shop. You get a credit balance equal to your subscription price. Each prompt costs whatever the underlying API call costs at retail rates. When the tab runs dry, the agent stops.
GitHub is explicit about this part: "Fallback experiences will no longer be available." No more silent downgrade to a discount model. You either pay or you stop.
A quick look at the numbers:
| Plan | Monthly price | Monthly AI Credits | Notes |
|---|---|---|---|
| Copilot Pro | $10 | $10 | $10 in retail tokens, no buffer |
| Copilot Pro+ | $39 | $39 | $39 in retail tokens |
| Copilot Business | $19/user | $19 (or $30 promo) | Promo lifts to $30 through Aug |
| Copilot Enterprise | $39/user | $39 (or $70 promo) | Promo lifts to $70 through Aug |
The "$X subscription gets you $X in credits" part is the line to stare at. The price of a Copilot subscription is now exactly equal to the value of the API tokens it bundles. There's no built-in discount on inference. The subscription is pricing itself as the wrapper, the UX, the convenience — not the model time.
Code completions and Next Edit suggestions still stay free across all plans. The meter only runs on the heavy stuff: chat, deep code review, agent sessions, tool calls.
Why this is happening (in honest terms)
Here's the part GitHub didn't say outright but is plain in the announcement.
The agent era ate the buffet model.
When Copilot was an autocomplete sidekick, the average premium request was tiny — a few hundred tokens to suggest the next line, maybe a thousand for an explanation. You could subsidize that all day on a $10 subscription.
Then we got agents. Long, multi-step sessions where the model reads twelve files, plans, tries, fails, retries, and finally writes a patch. A single agent run can chew through more tokens than a month of old-school autocomplete. GitHub's own framing is direct: Copilot has "evolved from an in-editor assistant into an agentic platform capable of running long, multi-step coding sessions."
There's no version of "unlimited" that survives that. Either everyone subsidizes the heavy users (and the price has to triple), or you meter the heavy users and let the light ones keep paying $10. GitHub picked the second option.
We've actually seen this movie before — about a month ago, when OpenAI introduced its $100/mo Pro plan with metered Codex. Anthropic's Pro plans use the same idea: subscription as a credit deposit, not an entitlement. GitHub holding out longer just means they tried hardest to keep the buffet alive.
Who wins, who loses
Let me try to be honest about both sides.
| Group | Old plan | After June 1 |
|---|---|---|
| Light users (autocomplete + occasional chat) | Plenty of room | Still plenty of room — completions stay free |
| Heavy chat users | Quietly downgraded mid-month | Burns through credits faster than you'd guess |
| Agent power users | Hit caps, kept going on cheap models | The expensive group now |
| Team admins | Had no real per-user budget knobs | Pooled credits + per-cost-center caps |
| Annual subscribers | PRU pricing | Grandfathered until renewal, then Copilot Free |
A few of those deserve a sentence.
The agent power users are the ones to call out. A single multi-file refactor with a frontier model can easily eat $1–$3 in retail tokens. Three of those a day, twenty days a month, and you're well past your $39 Pro+ allowance — which makes sense, because at retail token prices, $39 just isn't a lot of frontier-model time.
Team admins quietly got a present. The new admin budget controls let an org set ceilings per-user and per-cost-center, which the old PRU system couldn't really do. If you've ever had a senior engineer accidentally torch the team's premium quota with a single overnight agent loop, this is the fix.
Annual subscribers get a soft landing — current pricing holds until renewal, then it converts to Copilot Free unless someone actively re-subscribes. Set a calendar reminder, because "free" here means feature-limited, not a free version of your current plan.
The thing nobody is saying out loud
A subscription where the credit equals the subscription price is a vendor admitting that the inference is the product, and the wrapper is the bonus.
That's not a knock on GitHub — it's just where the industry has landed. Frontier-model inference costs real money. Anthropic, OpenAI, and now GitHub have all converged on the same answer: stop pretending a flat fee can cover unbounded compute, and bill people for what they actually use.
The honest framing would have been: "We're switching to metered pricing because agentic coding made flat-rate plans impossible." What we got was "continued sustainability concerns drive the shift toward actual usage alignment." Same thing, different vibe.
The reason this matters beyond Copilot is that the floor for "what an AI subscription means" is now visible. It's the underlying API cost. Anything you pay above that is for polish — UI, integrations, workflow, support. Useful thing to know when you're shopping plans across vendors.
What I'd actually do
A few practical moves before the meter goes live:
- On annual Pro and you mostly use completions? Do nothing. Completions stay free. You're already winning.
- On monthly with heavy agent use? Watch the preview billing GitHub is rolling out in early May. Get a real read on your token spend before the meter hits production.
- Running a team? The pooled credit feature is genuinely better than the old per-seat quota. Set per-cost-center caps now, not after someone burns through them.
- Living in chat? Try a cheaper model tier for routine questions. Save the frontier model for things where it actually matters. The fallback is gone — you're the fallback now.
The buffet was nice while it lasted. The meter is fairer, but it's also a meter. Adjust accordingly.
Sources
- GitHub Copilot is moving to usage-based billing — official announcement with pricing tiers, effective dates, and admin controls
- OpenAI's $100 Plan: Did They Copy Claude's Homework? — earlier post on how Anthropic and OpenAI converged on metered subscriptions
- Tracking the history of the now-deceased OpenAI Microsoft AGI clause — context on the same week's other AI-economics story