At 00:00 UTC today, GitHub Copilot stopped counting premium requests and started counting AI Credits. Same list price, different meter. If you haven’t opened your usage dashboard yet, you’re about to find out whether your particular flavor of “I use Copilot a lot” still fits inside your plan or whether you owe Microsoft a surprise this month.
I’ve been on the Pro+ plan since the preview rolled out in May, and the short version is this: light chat users won’t feel anything. Agent-mode users running Opus 4.7 on long tasks will feel it within a week. And for the first time since Copilot launched, the math actually makes some workloads cheaper to move to Cursor or Claude Code direct.
Here is the field guide I wish someone had handed me yesterday.
What actually changed at midnight
The Premium Request Unit (PRU) is gone. Each plan now ships with a dollar-denominated AI Credit budget — 1 credit equals $0.01 — that gets debited at the underlying model’s published API rate for every token you send and every token you receive. Cached input tokens are billed at the model’s cached rate, which on most providers is roughly a tenth of the fresh-input rate.
What isn’t metered, and this is the part GitHub is correctly putting in 60-point type on the announcement page:
- Code completions (the gray ghost text)
- Next edit suggestions
- Inline tab-complete on any plan
Those stay unlimited on every paid tier. If you only use Copilot for tab-complete, your bill is identical to last week and you can stop reading.
What is now metered:
- Chat in the editor
- Agent mode (any model, any length)
- Background tasks and cloud agents
- Model selection beyond the default
- Any premium model invocation (Opus 4.7, GPT-5.5, Gemini 3.5 Pro)
The default chat model on every plan is Haiku 4.7, which burns credits at roughly a tenth the rate of Opus. This matters more than the announcement post suggests.
Plan by plan, in plain English
GitHub kept the list prices identical, which is the bit that makes “no price change” technically true while completely changing the value calculation.
Pro — $10/mo, $10 in credits. Aimed at hobbyists and side-project developers. You get the base model lineup, no priority access. Realistically this covers maybe 150–200 chat prompts a month on Haiku, or about 30–50 short agent tasks. If you’ve ever opened agent mode and let it run a multi-file refactor, you’ve already burned through Pro’s monthly budget in one session.
Pro+ — $39/mo, $39 in credits. Premium model access (Opus, GPT-5.5, Gemini 3.5 Pro) and priority queueing. This is the plan for the individual professional who actually relies on Copilot. You get roughly 4× the budget of Pro, but premium models burn credits 8–12× faster, so the net is “you can use Opus, but not all day.”
Business — $19/seat, $19 in credits. Team management, audit log, IP indemnification, but only the base model lineup unless an admin explicitly enables premium models for the org. Most Business seats will hit their cap mid-month if engineers freely use chat. Admins can buy top-up credits at $0.01 each, billed to the org.
Enterprise — $39/seat, $39 in credits. Same as Business plus SSO, audit retention, and premium models on by default. The seat price stayed the same, which is fine if your engineers are casual chat users and a bad joke if you have a platform team running scheduled background agents.
Max — custom pricing, highest credit allowance, priority model access. The new top SKU, aimed at organizations that pre-commit annual credit budgets and want priority during peak. Microsoft won’t quote it without a sales call, which tells you everything about the target buyer.
Annual subscribers who locked in before June 1 stay on legacy PRU billing until their renewal date, then convert to AI Credits prorated for the remainder of the term. If you paid annually in May 2026 specifically to dodge this, you bought yourself eleven months.
The math by model
This is the part the announcement glosses over. Credit burn depends entirely on which model you call and how big your context is. Rough numbers, based on each provider’s published rates as of this week:
Claude Opus 4.7 — input around $15 per million tokens, output around $75. A single agent-mode PR with a 40K context window and 8K of output runs roughly 60–100 credits. Cached tokens drop the input bill by about 90%, but only if your prefix is stable across calls. One long agent task per day, every weekday, lands you around 1,200–2,000 credits a month — well past the Pro+ cap.
GPT-5.5 — input around $10 per million, output around $40. Roughly 40–70 credits per equivalent PR-sized task. Cheaper than Opus on raw tokens, but in my testing it needs more turns to land the same change, so the per-PR cost ends up in the same neighborhood.
Gemini 3.5 Pro — input around $7 per million, output around $30. Cheapest of the three frontier models per token. Around 30–50 credits per PR-sized task. Google has been undercutting on token price all year and Copilot’s new model picker exposes that.
Claude Haiku 4.7 — input around $1 per million, output around $5. Around 3–8 credits per equivalent task. If you can do your work with Haiku, AI Credits are mostly invisible.
I ran a deliberate experiment yesterday: same six-step refactor across all four models in agent mode, with the same starting prompt. Opus finished it in two turns and burned 78 credits. GPT-5.5 took four turns, 61 credits. Gemini took three turns, 44 credits. Haiku took six turns and got lost on step five, but the credits used were 11. Your mileage varies by task type, but for code-heavy multi-file work Opus is still the winner per dollar despite the higher token rate, because it needs fewer turns. The Haiku result is the honest caveat — if your task is fuzzy or genuinely needs reasoning, the cheap model fails and you pay twice.
Three real scenarios
The press coverage is full of abstract “X credits per Y” tables. Here is what actually happens to three real workloads.
The light chat user. You ask Copilot maybe 5–10 questions a day, mostly “explain this regex” or “rewrite this SQL.” Default model is Haiku, average 4 credits per exchange, 30 exchanges a month = 120 credits. Pro’s $10 budget covers you with room to spare. No change.
The heavy agent-mode user. You let agent mode run 10–20 multi-file tasks a day on Opus 4.7. Average 80 credits per task. 15 tasks × 20 working days = roughly 24,000 credits, or $240/month. Pro+ at $39 covers about one work-day of this volume. You will either downshift to Haiku for most tasks, top up credits at full retail, or move the workload off Copilot entirely.
The platform team running background agents. Your CI runs a Copilot background task on every PR — security review, doc generation, test scaffolding. Average task uses 50K input + 5K output tokens, mixed cached + fresh. With 200 PRs a week and Opus selected, that’s roughly 4,000 credits/week, or 16,000/month. Multiply by however many seats you assign as the “agent runner.” This is where Enterprise pricing falls apart fastest, and where most ops teams will quietly pin the model to Haiku or Gemini.
The migration traps
A few things that bit me or my team this week and aren’t on the official “what’s new” page:
Cached tokens are billed differently in agent mode than in chat. Long agent runs that re-send the same system prompt across many turns benefit from caching, but only after the second turn — the first call to a new model session pays full input rate. If your tasks are short and you switch models mid-session, you get the worst of both worlds.
Per-org budget controls only catch overages, not project them. The org admin dashboard shows current spend and recent burn rate. There’s no forecasting, no “you’ll hit your cap on day 23 at current pace” warning. Build that yourself with the usage API if you care.
Buying additional credits requires admin approval on Business and Enterprise. Individual seats can’t self-serve a top-up, even at full retail. This is intentional. It also means that on day 28 of the month, your senior engineer might be tab-completing in HTML while waiting for Slack approval to buy $20 of credits.
Default model selection in VS Code didn’t migrate cleanly. A lot of people had Opus pinned in their settings from the preview. As of this morning, fresh VS Code installs default to Haiku. Pinned settings carry over, but if you spin up a new dev container or fresh laptop, watch for Haiku quietly running in agent mode while you wonder why the suggestions are worse.
When another tool is cheaper
I’m not going to pretend Cursor or Claude Code is universally better — Copilot has the deepest VS Code integration of any of them, the IP indemnification is real, and the enterprise audit story is more mature. But the pricing math changes today.
Cursor at $20/mo for the Pro tier still has the most generous “we’ll figure out the meter later” stance in the market. If you live in agent mode and your workload looks like the heavy-user scenario above, Cursor’s flat Pro plan plus targeted usage purchases comes in cheaper than Pro+ on Copilot for most months. Cursor’s own Pro+ tier launched in March 2026 at $40, but the base $20 plan covers more agent invocations than Copilot’s Pro+ on equivalent models. The catch: Cursor’s quota is opaque, and they reserve the right to throttle anyone whose usage looks like an outlier.
Claude Code billed directly through Anthropic, using your own API key, removes the middleman markup entirely. You pay published Anthropic rates with no platform fee. For the heavy-user scenario, that’s roughly the same dollar amount as Copilot Pro+ would charge, but you get unlimited Opus access bounded only by your actual API budget. The trade-off is no VS Code chat panel, no inline completions — Claude Code is a terminal-first agent tool, not a Copilot replacement.
JetBrains AI Assistant at $20/mo includes a fixed quota of premium model calls plus unlimited completions on JetBrains’ own model. If you live in IntelliJ or PyCharm, this has always been the underrated choice, and the new Copilot pricing makes the gap wider.
Codeium / Windsurf stays free for individuals on their own model, with paid tiers for the Cascade agent and frontier model access. Windsurf’s $15/mo Pro plan is the cheapest entry into agent-mode coding from a major vendor right now.
AWS Kiro and Google Antigravity 2 are both viable if you already live in those clouds and your org has committed credits to burn. Antigravity 2’s free tier specifically was expanded in May to compete with this very announcement.
Decision tree, short version
- Tab-complete only: stay on Pro. Nothing changed.
- Light chat user: stay on Pro. Watch one month, then decide.
- Heavy chat user, no agent: stay on Pro+, default to Haiku, escalate to Opus only on hard problems.
- Heavy agent user: try Cursor for two weeks against your real workload. Compare the bills. Move if the math is clear.
- Platform team running scheduled agents: pin the model to Haiku or Gemini, or move that workload to Claude Code with your own API budget. Don’t run Opus in unattended CI.
- Enterprise with annual lock-in: you have eleven months. Use them to actually instrument usage and pick the next contract on data, not on what your dev advocate likes.
What to watch next
Microsoft will react to the headlines. My guess is a quiet Pro+ credit bump within 90 days, probably bundled with whatever model lineup ships at GitHub Universe in October. The September promotional credit allotment for Business and Enterprise is the other cliff to watch — that ends and a lot of orgs will discover their actual sustained-usage cost for the first time.
The thing nobody is talking about loudly enough is that AI Credits make every IDE assistant directly comparable to every other one for the first time. Token in, token out, dollar denominated. That’s good for buyers and uncomfortable for the vendors who were quietly subsidizing power users with the flat-rate model.
Pick one workload you actually do every day, run it through agent mode on each of the three frontier models, and watch the credits tick. That ten-minute experiment will tell you more about your real Copilot bill in July than any pricing page will.