Update · April 27, 2026 Copilot is moving to usage-based billing on June 1, 2026. PRUs are being replaced by GitHub AI Credits (1 credit = $0.01 USD), billed by token at published API rates. Plan prices are unchanged. The "hypothetical" token-billing scenario in this story is now confirmed reality. Announcement · Pooling for orgs · Per-token pricing.
A Copilot pricing story · confirmed June 1, 2026

GitHub is replacing PRU billing with AI Credits. Who actually wins?

Copilot meters most models in premium request units — fixed-price chats with a per-model multiplier. Direct API providers meter in tokens. The two systems disagree by an order of magnitude in opposite directions, and which way they disagree depends entirely on how you use Copilot. On June 1, 2026 GitHub is replacing PRUs with GitHub AI Credits (1 credit = $0.01 USD, billed per token at published API rates). Plan prices are unchanged. This page models chat, agent mode, Copilot CLI, sub-agents and the cloud agent across both systems — and adds a dedicated pooling visualization for the Business and Enterprise plans, which now share a single credit pool across the org.

Heaviest PRU markup (chat only)
Opus 4.7 PRU vs simple-chat tokens — disappears under AI Credits
Heaviest PRU subsidy (cloud agent)
Same model, large cloud-agent task — bills as tokens after Jun 1
Workflow modes covered
7
From simple chat to large cloud-agent runs
Business pool · 100 seats
190k credits
Pooled across the org · 300k during Jun–Aug 2026 promo

1The question, answered on April 27

On a paid Copilot plan today, a chat with GPT-4o costs $0, a chat with Claude Opus 4.7 costs $0.30, and the actual provider token cost of those two conversations is roughly $0.03 and $0.07. PRU and tokens disagree by an order of magnitude in opposite directions. On April 27, 2026 GitHub announced it is flipping the switch — from premium request units to GitHub AI Credits, metered per token at published API rates — on June 1, 2026. So what actually moves?

Two-part answer. For chat: ultra-premium gets ~4× cheaper, the included tier disappears, and Gemini wins the mid-tier. For agentic workflows: the math reverses completely — PRU's "tool calls don't count" rule was worth $300–$2,000/month for agent-heavy users, and unwinding it dwarfs every other change. For Business and Enterprise: credits are pooled across the org — see section 9.

2The five PRU tiers, and the token costs that explain them

Copilot collapses 18 models into five pricing tiers. Each tier's multiplier is a deliberate choice — sometimes a fair reflection of the model's underlying token cost, sometimes a strategic subsidy or markup. Here's what each tier actually buys you and why GitHub almost certainly priced it that way.

Read this as a system, not a list. The 7.5× ultra-premium tier funds the 0× included tier. The 1.25× high-premium tier is consistently subsidized at chat sizes and at agent sizes — a deliberate "use Opus, don't be afraid" signal. Standard Premium (1×) is the only tier that's roughly cost-neutral against tokens.

3Every model placed on the multiplier ladder

Eighteen models, five tiers. Hover any bar to see the multiplier, the per-chat PRU charge, and the underlying token cost for that model.

Copilot premium-request multiplier by model
Higher = more PRU per chat. Bars colored by tier. Multiplier↔list-price-token-cost correlation across these models is r ≈ 0.80 — real but loose.

4The hidden subsidy — and why the chat picture is wrong

Subtracting the actual provider-token cost of a typical chat from the PRU charge reveals who is over- and under-charged. The first toggle below is the original analysis (4K-in / 2K-out simple chat). The second toggle re-runs the same comparison at agent-mode token volumes (60K in / 15K out). Watch the ultra-premium bars flip sign.

Read this as a user-perceived subsidy, not a P&L subsidy. Every "subsidy" figure on this page benchmarks the PRU charge against provider list price (Anthropic, OpenAI, Google public API rates). GitHub's actual cost-to-serve is materially lower because of negotiated commit rates, the Microsoft↔OpenAI revenue-share, Azure-hosted inference, and provider-side prompt caching / batch discounts. The relative ordering of subsidies across tiers is preserved, but the absolute GitHub→user transfer is smaller than the bars below imply — and on paid plans the included tier's "zero marginal cost" is bundled into the seat fee, not paid for by GitHub out of pocket.
PRU price minus provider list-price token cost · per task
Right (red) = PRU charges more than list price · Left (green) = PRU charges less than list price (user-perceived subsidy). GitHub's negotiated cost-to-serve is lower than list, so absolute green-side magnitudes overstate the P&L cost to GitHub.
The "ultra-premium overcharges users" finding only holds for short chats. The moment a user prompt triggers any kind of agentic loop — workspace context, tool calls, tests, sub-agents — the high and ultra tiers flip from the red side to the green side. PRU subsidizes users of those tiers in the modes those tiers were designed for. Note also that the 1.25×→7.5× jump is a 6× price step against a 1× step in list-price token cost — that gap reflects reasoning/adaptive-thinking output volume, promotional launch pricing (Opus 4.7's 7.5× rate expires April 30, 2026), and deliberate cross-subsidy of the free and agentic tiers, not raw input/output rates.

5Why "chat" is the wrong unit

Copilot in 2026 is not one product — it's at least seven workflow modes spanning nearly three orders of magnitude in token consumption. The 4K-in / 2K-out "typical chat" figure that the original PRU vs. token comparison rested on describes only the leftmost column of this picture.

Token volume by workflow mode (log scale)
One bar = one full task or session. The same prompt costs the same in PRU; under tokens it scales with the bar.
The "tool calls don't count" rule. The Copilot billing docs are explicit: "only the prompts you send count as premium requests; actions Copilot takes autonomously to complete your task, such as tool calls, do not." That single sentence is the most valuable subsidy in the entire Copilot pricing system, and it grows non-linearly with workflow autonomy.

6Cost vs. SWE-bench performance

The headline trade-off: how much does each model cost per chat, and how does that map to its SWE-bench bash-only resolve rate? Toggle between PRU pricing and token pricing to watch the frontier reshuffle.

SWE-bench resolve rate vs. cost per typical chat
Up-and-to-the-left is best. Bubble color = provider.

7Run your own numbers — with workflow mode

Pick a model, a workflow mode, and a monthly volume. The chart compares your bill today (PRU) versus what it becomes on June 1, 2026 (per-token / AI Credits — where the workflow mode dominates).

PRU monthly (today)
$0
AI Credits monthly (Jun 1)
$0
Delta (PRU − Credits)
$0

8Winners and losers under the new AI Credits billing

Each row models a real user profile under the actual workflow mix that profile uses — not a chat-only assumption. Bars show the monthly-cost range under per-token AI Credits pricing (purple) versus today's PRU bill (blue).

Monthly Copilot cost by user profile · today (PRU) vs. June 1, 2026 (AI Credits)
Blue bar = today's PRU charge. Purple range = AI-Credits monthly bill (low/high). Right-of-blue = the user pays more after Jun 1.

9Pooling for Business & Enterprise

Individual Pro and Pro+ plans get their AI Credit allowance per seat, burnable only by that user. Copilot Business and Enterprise pool every seat's allowance into one organization-wide bucket — heavy users draw from the same pool as light users, smoothing out the per-seat cliff that worried agent-heavy teams. During the introductory period (June 1 – September 1, 2026) the Business and Enterprise allowances are increased to 3,000 and 7,000 credits per seat respectively.

Plan allowances at a glance. Pro $10/mo → 1,000c/seat · Pro+ $39/mo → 3,900c/seat · Business $19/mo → 1,900c/seat (3,000c/seat during Jun–Aug 2026, pooled) · Enterprise $39/mo → 3,900c/seat (7,000c/seat during Jun–Aug 2026, pooled). Code completions and Next Edit Suggestions remain unmetered.
Pool300,000 credits = Subsidy$3,000/mo Per seat3,000c
Four real fleet shapes — does the pool cover the bill?
Each card models 100% of the org's monthly token cost across one of four usage shapes. Green = light-user consumption that fits inside the pool. Red = power-user consumption. The vertical line is your pooled monthly allowance. Striped overflow = paid usage on top of the subscription.
Light-user credits used Power-user credits used Headroom in pool Overage (paid)
Why pooling matters in practice. A single agent-heavy engineer can easily burn 10,000+ credits/month. Without pooling that engineer either hits a hard wall mid-month or the org has to pre-budget overage seat-by-seat. With pooling the same engineer just dips into the 99 dormant seats' allowances — and the IT admin sees one number, one budget control, one bill. Adding licenses mid-cycle increases the pool immediately. Removing them takes effect at the next billing cycle.
Caveats to read carefully. There is no automatic fallback to a cheaper model when credits run out — the admin must either set an overage budget, or affected users get blocked. Copilot code review charges both AI Credits and GitHub Actions minutes. Annual Pro/Pro+ subscribers stay on PRU, but per-model multipliers are rebased significantly (Opus 4.6 from 3× to 27×, Sonnet 4.5 from 1× to 6×). See the analysis doc for the full multiplier table.

10The verdict, rewritten

The PRU model is a deliberate cross-subsidy: ultra-premium chat overcharges fund the zero-cost included tier, and PRU's "tool calls don't count" rule funds every agentic workflow on every tier. The June 1 transition unwinds both subsidies — and the second one is much, much larger than the first. Pooling is GitHub's answer for the second.

✅ Chat-only ultra-premium users save

Opus 4.7 / GPT-5.5 in plain chat mode drop from $0.30 to ~$0.07–0.08 per turn. A 200-prompt user saves ~$44/month after Jun 1.

⚠️ The included tier disappears

GPT-5 mini, GPT-4.1, GPT-4o all start metering at $0.005–$0.030/chat. Casual users on paid plans pay something for the first time.

⚠️ Agent mode users lose hundreds — unless pooled

One IDE agent task burns 5–10× a chat in tokens. Heavy individual users pay $200–600/mo extra. On Business/Enterprise, the pool absorbs much of it.

⚠️ The cloud agent becomes brutal

One large cloud-agent task = $10+ of tokens at provider list price, vs. $0.30 in PRU. Five tasks/day on Opus = ~$1,500/month extra without pooling.

🏷️ Cheap models burn more tokens

Gemini 3 Flash at $0.50/MTok solves the same SWE-bench task as Opus at 2.5× the token count. Token pricing rewards iteration efficiency, not just rate.

💾 Caching becomes the budget lever

Anthropic / Gemini cached input is 90% off. A stable 10K-token system prompt becomes effectively free under AI Credits — PRU never rewarded this.

🤝 Pooling is the org's safety net

For Business/Enterprise, allowances pool across the org. The introductory Jun–Aug period adds ~58% (Business) and ~80% (Enterprise) extra credits per seat.

📅 Annual Pro/Pro+ multipliers are rebased

Annual subscribers don't move to AI Credits but their PRU multipliers jump (Opus 4.6 3×→27×, Sonnet 4.5 1×→6×, GPT-5.4 mini 0.33×→6×) — economically equivalent.

Bottom line. Two subsidies hold up Copilot's pricing today. The ultra-premium overcharge funds the included tier. The "tool calls don't count" rule is the much bigger subsidy — it funds every autonomous workflow on every tier. The June 1, 2026 transition exposes both to per-token reality. Casual chat users will notice; agent-heavy individual users will feel it like a tax; Business and Enterprise organizations get pooling and an introductory bonus to cushion the move.
Now do it with your numbers

Analyse your own Copilot usage

Everything above is industry-wide modelling. If your enterprise has access to the Copilot metrics export, drop the NDJSON file into the upload page and get the same kinds of charts — feature mix, model mix, language acceptance, top users, CLI token burn — for your seats. Parsing runs entirely in your browser.

Upload your Copilot usage data →

🔒 Parsed locally in your browser. Your data never leaves this page.

Take this with you

Export the full data story — narrative, tables, and every D3 visualization — as a single PDF for offline reading or sharing.

Generates a multi-page PDF of this entire page. Large pages may take 10–30 seconds.