GitHub Copilot Switches to Usage-Based Billing on June 1. The Token Tab Came Due.

title: GitHub Copilot Switches to Usage-Based Billing on June 1. The Token Tab Came Due. description: Microsoft announced Copilot Pro is still $10 per month, but $10 now buys $10 in AI credits, and a single Opus agent session can consume that. Here is what your real bill looks like. tags: ai, webdev, devtools, productivity

The receipt landed in my GitHub inbox on a Monday morning.

Subject: important changes to your Copilot plan.

Body: starting June 1, every Copilot subscription pays a metered tab, not a flat fee. The Pro tier still costs ten dollars, but ten dollars now buys you ten dollars in AI credits, and a single agent session against Claude Opus can burn that in one run.

I published a piece last week called The Token Tab: A Developer's Audit of the AI Hype Stack. The thesis was that AI tools advertise layer 1 (the wrapper) and layer 4 (the hardware), then quietly bill you for layer 2 (orchestration) and layer 3 (inference). I did not expect the largest AI dev tool on the planet to prove the thesis four days later with a public billing change. But here we are.

Microsoft is not killing Copilot. They are unbundling it.

The four things your ten dollars actually pays for now

Before June 1 the bundle was simple. You paid ten dollars and got everything: tab completions, chat, agent mode, occasional Opus access. Microsoft ate the cost difference between users who barely opened the chat panel and users who ran agent mode all day.

Starting June 1, the bundle splits. Your ten dollars now resolves to four distinct things. Two are still free. Two run on a meter.

One AI credit equals one cent. Pro includes $10 in credits per month. Pro+ includes $39. Business includes $19, with a $30 promotional credit through August. Enterprise includes $39, with a $70 promotional credit through August. Unused credits do not roll over.

Layer 1: completions and Next Edit

This is the load-bearing piece for most developers. The autocomplete on the right side of your function definition is unchanged. Microsoft keeps it included because it is the entry-level habit. Meter it and people churn out instantly.

If you only use Copilot for autocomplete and Next Edit, your bill stays at ten dollars and nothing about June 1 applies to you. Stop reading and ship something.

Layer 2: chat is where the credit math begins

Chat is metered per token. Microsoft uses the published API rates for each model. Claude Sonnet 4.6, the default chat model on most plans, costs $3 per million input tokens and $15 per million output tokens.

Here is what a modest day looks like at those rates. A developer asking ten chat questions, averaging 10K tokens of context loaded and 2K tokens generated:

• Input: 100K tokens at $3 per million = $0.30, or 30 credits
• Output: 20K tokens at $15 per million = $0.30, or 30 credits
• Daily total: 60 credits, or $0.60

A working month at that rate burns $18, against a Pro plan budget of $10. You hit zero on day seventeen.

If you mostly skim docs and ask short questions, that is fine. If you treat Copilot Chat as a coding partner, you run a deficit by week three.

Layer 3: agent mode and the multiplier trap

Agent mode is where the math stops being modest. An agent session loads more context, often 50K to 200K tokens. It runs a tool loop with multiple model invocations. It frequently uses premium models like Claude Opus 4.7, priced at $5 per million input and $25 per million output.

A representative agent task against Opus 4.7:

initial context load:        80K in           = $0.40
five tool iterations:        150K in / 40K out = $1.75
final synthesis:             50K in / 10K out  = $0.50
                                                ------
                                                $2.65

Single session, conservative numbers: 265 credits. Your $10 Pro budget covers roughly four of these per month before overage kicks in.

Annual subscribers staying on the older premium-request model are not safe either. Microsoft is increasing model multipliers on June 1. Opus 4.7 is reportedly priced at 7.5 requests per query. A user with 300 monthly premium requests gets 40 effective Opus prompts.

Layer 4: the fallback model that just disappeared

This is the change buried in the announcement that nobody is talking about.

Before, when you exhausted your premium request quota, Copilot would fall back to a cheaper model so chat kept working. The April 27 announcement says: "Fallback experiences will no longer be available."

Translation: when your credits run out, Copilot stops responding. No degraded mode. No silent downgrade. You either top up or you wait until next month.

For developers who treat Copilot as ambient infrastructure, that is the actual breaking change. Not the price. The behavior.

I priced 30 days of my own usage at the new rates

I pulled my Copilot dashboard for the last 30 days and applied the June 1 rate card. Real numbers, one developer, me:

I am on Copilot Pro+ at $39 per month, which now includes $39 in credits. Under the new model, my last month would have run $39 plus $13 in overage, or $52 total. Up from a flat $39.

That is a 33 percent increase for the same workflow. Not 10x, not catastrophe, but a real and uncapped surcharge for anyone who runs agent mode regularly.

For a Copilot Business team of ten doing similar work, the math compounds: $520 plus $130 in overage against $190 flat. The promotional $30 credit per seat through August softens the impact. After September 1, the meter runs at full rate.

Three signals to watch between now and June 1

The preview bill launches in early May. That is the first signal. Anyone who ignores it and discovers their actual usage on June 1 will have a bad week.

Watch the rate card itself. GitHub's pricing page will update as model providers adjust their wholesale rates. Anthropic and OpenAI both raised inference prices once in 2026. A second round before September is plausible.

Watch what happens to the IDE plugin. If Microsoft starts metering Next Edit suggestions next, the $10 Pro plan becomes nominal. The free completions tier is the entire reason Copilot has 20 million users. If that tier moves to credit consumption, the calculation changes again.

Five things to verify before the auto-migrate

This week, before May 1:

1. Open your Copilot billing dashboard and pull last month's usage breakdown. Numbers in hand make every other decision easier.
2. Decide whether your annual plan is worth converting early for prorated credits or riding out at the current premium-request rate. The math depends on agent use.
3. If you run agent mode against Opus, audit which tasks actually need Opus and which run fine on Sonnet 4.6 or GPT-4.1. Sonnet costs $3 per million input vs Opus at $5, and the output gap is wider.
4. Set a credit budget in your billing controls. Without one, June will surprise you.
5. Test the API direct path. Anthropic and OpenAI ship cheaper rates than the Copilot middleman markup for power users. Cursor, Aider, and direct API integrations are no longer obvious losers on price.

Closer

Last week I argued the AI hype stack hides its costs in layers 2 and 3. This week the largest AI dev tool on the planet ripped the wallpaper off and printed the meter on the front of the box.

Whether that is good or bad depends on whether you are a heavy user paying a discount today or a light user subsidizing the heavy ones. My guess: Microsoft will discover the heavy users churn faster than the light users replace them. We will know by Q3.

What does your last 30 days of Copilot usage look like at the new rates? Drop your number in the comments. I am collecting real profiles for a follow-up.

Companion piece on the supply-chain side of the AI tooling layer: An AI Tool Had OAuth to Their Whole Google Workspace. Then Vercel Got Breached. (Replace with the published URL once the second article is live.)