GitHub’s fixed-rate era for premium inference ended June 1, 2026. GitHub announced on April 271 that Premium Request Units (PRUs) across all Copilot plans were replaced by GitHub AI Credits billed against actual token consumption (input, output, and cached) at per-model API rates. Monthly plan prices are unchanged. The per-request subsidies are not.
What Changed and When
The PRU system gave every Copilot request a fixed cost expressed as a multiplier of a base rate. According to GitHub’s premium request documentation2, now labeled legacy, the launch multipliers were Claude Opus 4.7 and GPT-5.5 at 7.5x, Claude Opus 4.5 and 4.6 at 3x, Claude Sonnet 4.6 at 1x, and Claude Opus 4.6 in fast mode (preview) at 30x. Those rates escalated for the annual subscribers who stayed on PRU billing past June 1: Opus 4.7 reached 27x and Sonnet 4.6 9x. Anthropic shipped Claude Opus 4.8 on May 28, and GitHub gave it a 15x premium-request multiplier for the few days until the June 1 cutover. (GitHub Blog. “Claude Opus 4.8 is generally available for GitHub Copilot.” May 28, 2026) The model is now listed in GitHub’s rate card4 at the same $5/$25 per million token price as Opus 4.7. That multiplier table expired June 1.
Since June 1, a request to Opus 4.7 or the newer Opus 4.8 costs whatever its token footprint resolves to against GitHub’s published API rates for that model. The 7.5x multiplier was a fixed cost regardless of context window size. A request that sends 200k tokens of context to an Opus flagship model now costs proportionally more than a 10k-token request. The PRU system couldn’t express that distinction.
GitHub’s per-model rate card is now live on Models and Pricing for GitHub Copilot. Each token is priced based on the model used, and the total is converted to credits at 1 AI credit = $0.01 USD4. The published rates went into effect at the June 1 cutover.
The New Credit Economics by Plan
Monthly base prices are unchanged, and base credits still match each plan’s price dollar-for-dollar. On May 12, though, GitHub added a variable “flex” allotment on top of the individual plans and introduced a new Copilot Max tier; the flex portion is one GitHub says it will adjust “as the economics of AI evolve.” (GitHub Blog. “Flex allotments in Pro and Pro+, and a new Max plan.” May 12, 2026) Current included usage by plan:
| Plan | Price | Monthly included usage | Promotional Credits (June-August 2026) |
|---|---|---|---|
| Copilot Pro | $10/month | 1,500 credits ($15: $10 base + $5 flex) | N/A |
| Copilot Pro+ | $39/month | 7,000 credits ($70: $39 base + $31 flex) | N/A |
| Copilot Max (new) | $100/month | 20,000 credits ($200: $100 base + $100 flex) | N/A |
| Copilot Business | $19/user/month | 1,900 credits ($19) | 3,000 credits/user ($30) |
| Copilot Enterprise | $39/user/month | 3,900 credits ($39) | 7,000 credits/user ($70) |
Business and Enterprise plans receive bonus credits through August 20261. That buffer disappears September 1.
Credits don’t roll over. Unused credits in a billing cycle reset at the next renewal.
Code completions and Next Edit Suggestions remain included in all plans and consume no credits. The cost exposure is in chat, agent mode, and code review.
How Agent Workflows Get Repriced
The structural change hits hardest in agentic contexts. Under PRUs, a team running an agent loop with multiple Opus 4.7 calls paid a fixed 7.5x per request regardless of context size. Token-metered billing removes that ceiling. An agent that accumulates tool-call outputs across a long session and re-sends the full context window on each turn now pays for the accumulation.
Groundy covered the 7.5x→15x→27x multiplier arc and the April 20 plan changes as a separate story; the June 1 change retired the multiplier table entirely rather than adjusting it. The economic bet the old system implicitly made was that average context sizes would be small enough to absorb. Evidently they weren’t. For broader context on how this fits in the wider AI coding tool landscape, our Copilot vs Cursor vs Claude Code 2026 showdown tracks the three tools’ divergent business models through this transition.
GitHub also introduced session and weekly limits based on token consumption that operate separately from premium request entitlements and can block usage even when request budgets haven’t been hit. Teams relying on agent mode need to account for both the per-token cost and the throughput ceiling.
The Hacker News discussion of the announcement3 drew considerable attention, with threads covering cost shock on the Pro+ tier, the absence of rollover, and comparisons to direct OpenRouter access at API rates without the subscription overhead.
Annual Subscriber Gotchas
Annual Pro and Pro+ subscribers didn’t convert to credit billing on June 1. According to GitHub’s announcement, they stay on PRU-based pricing until their plan expires, but their model multipliers increased on June 1 regardless. At renewal, they drop to Copilot Free unless they convert early, in which case they receive prorated credits.
An annual subscriber who renewed in January 2026 keeps PRU billing until January 2027 but pays higher multipliers from June 1 onward. At renewal they either convert before the expiration date or land on the free tier with no premium model access.
New sign-ups for Copilot Pro, Pro+, and Student plans were paused April 20, ahead of the billing change. Opus models were removed from Pro plans at the same time; Opus 4.7 remains in Pro+, joined by Opus 4.8 on May 28.
The Hidden Code-Review Tax
Copilot code review picked up a second charge on June 1. GitHub’s changelog states that code review on private repositories will consume both AI Credits and GitHub Actions minutes starting June 1. Public repositories are unchanged.
Teams running automated review on private repos (particularly those with high PR volume or large changesets) face a compound cost that doesn’t appear in the monthly plan pricing. The Actions-minute component charges against each team’s existing Actions quota, separately from the AI Credits component.
What to Audit After the June 1 Cutover
The credit amounts included in monthly plans are modest relative to heavy agentic usage. The Pro+ monthly allocation covers materially different workloads depending on which models and context sizes are in use. With the rate card now fixed4, exact projections are possible.
Four areas to audit now that token billing is live:
Context window hygiene. Agent loops that send full conversation history on each turn inflate token counts per call. Audit tool-call output accumulation and context pruning strategies. What looked cheap at a fixed 7.5x per request may be expensive at per-token billing.
Model selection. Under PRUs, choosing between Opus 4.7 (7.5x) and Sonnet 4.6 (1x) was a fixed cost tradeoff. Under token-metered billing, that tradeoff becomes continuous and depends on actual input/output size per request.
Code review on private repos. Teams that added automated Copilot code review without tracking Actions-minute consumption should audit expected PR volume now that the dual charge is active.
Annual subscription dates. Subscribers renewing after June 1 face higher multipliers before their plan expires. Those close to renewal should decide whether to convert early for prorated credits or accept the Free tier drop.
The promotional credit buffer for Business and Enterprise runs through August 2026. September is when steady-state costs become visible.
Frequently Asked Questions
How long could a heavy Copilot Pro+ user realistically run agent mode before exhausting the monthly credit?
Hacker News commenters reacting to the original announcement noted that Pro+‘s monthly AI Credit allocation could be burned through in roughly an hour of intensive agentic coding with long context windows and repeated Opus calls. The May 12 flex allotment lifted Pro+ to about $70 of included usage, which extends that runway but leaves the buffer a safety net for light usage rather than a workable floor for power users.
How does the total cost of Copilot under token billing differ from buying model access through OpenRouter?
OpenRouter sells API access at published per-model rates without a subscription wrapper, whereas Copilot bundles subscription fees plus metered credits. Discussion in the community threads flagged OpenRouter as a direct escape hatch: teams doing mostly agentic work may pay less by routing model calls through OpenRouter while keeping only a minimal Copilot subscription for IDE completions, which remain included and credit-free.
Do the new token-based session and weekly limits apply to Copilot Business and Enterprise, or only to individual Pro plans?
GitHub introduced those throughput ceilings on April 20 as part of the individual-plan changes, and the announcement tied them specifically to Pro, Pro+, and Student tiers. Business and Enterprise plans are not explicitly mentioned as receiving the same session or weekly caps, though they remain subject to the AI Credit budget itself.
What is the cost risk for teams running automated Copilot code review on private monorepos with high PR volume?
The compound charge is steeper than the headline suggests because large changesets in monorepos trigger longer review jobs, consuming more GitHub Actions minutes per PR in addition to the AI Credits for the model call. Teams that previously treated Copilot review as a fixed subscription feature now face a variable cost scaled by both PR frequency and changeset size, with no promotional buffer covering the Actions-minute side of the bill.
When will teams be able to calculate exact breakeven points for their agent workflows?
GitHub launched the preview-bill flow in early May 2026 and now publishes per-model rates on the Models and Pricing reference page. Exact cost modeling for high-volume agent loops became possible at the June 1 cutover, with rates now fixed4 unless GitHub announces adjustments.