GitHub Universe 2026: How the Copilot Agent Changes Code Review, Throughput, and AI Coding Cost

The conference has not started, and the review bill already has. GitHub Universe 2026 is set for October 28–29. The sections below follow code review, developer throughput, and the cost of AI coding, using only what GitHub had written into docs and the changelog by October 10, 2026, plus the settings a team can check before the opening keynote.

What was already documented before opening day

On June 4, 2026, GitHub confirmed Universe would return to Fort Mason Center in San Francisco on October 28–29, in person and online, under the line “All together now, in the agentic era.” The schedule published on August 13 already lists “Stop prompting, start delegating” and a session on how the Copilot coding harness is evaluated across models. Those are session descriptions, not release notes. (GitHub’s event announcement, schedule guide)

The earlier watchlist is in our GitHub Universe 2026 preview. That piece is about what to look for. This one answers a narrower question: before the keynote, how has the Copilot agent already changed review, delivery pace, and the bill.

Code review: from a comment bot to a repo-aware agent

Copilot code review no longer stops at a few lines of diff. The docs say it uses agentic capabilities for full-repository context, and it can hand suggestions to the Copilot cloud agent, which opens a follow-up pull request with the fixes applied. That handoff is still a public preview. Context gathering and tool use run on GitHub Actions. If Actions is unavailable, a review is still posted, without those agentic extras. (Copilot code review docs)

Effort has two levels. Lite targets common bugs, security issues, and style, and returns faster. Balanced routes the pull request to a higher-reasoning model for complex logic, security-sensitive changes, and cross-service work. It uses more AI Credits and may use slightly more Actions minutes. GitHub’s August 28 changelog says that from September 28, 2026, an organization or repository left on Default uses Balanced. Keeping Lite means selecting Lite explicitly. (August 28 policy and billing update)

Those dollar figures are GitHub’s typical estimates and exclude GitHub Actions minutes. Larger pull requests and longer custom instructions usually cost more, and the range can move as models change.

A review can also read repository agent skills, custom instructions, and MCP servers. The GitHub MCP server and Playwright MCP server are on by default. Instructions and skills are read from the head branch, so a team can try a new review rule in the same pull request without merging it first. The Models page in organization settings controls Copilot Chat only. It does not change the review model. GitHub says switching that model would undermine review consistency.

Every review includes an approval assessment in the overview comment. By default it does not count toward required approvals. Copilot approvals, still a public preview, count like a teammate’s approval only after the repository, organization, and enterprise all enable them. A new push dismisses that approval. GitHub also says Copilot is not guaranteed to find every problem, so a human still has to read the result.

Two billing switches added on October 8

The October 8, 2026 changelog gave organization admins two controls. First, a review requested by a member who already has a Copilot license is billed to that member’s entitlement by default, and the review fails when the quota is gone. An organization owner can bill the organization that owns the repository instead, which requires paid AI Credit usage; a budget is optional. Second, admins can require that only a license issued by the organization or enterprise may trigger a review, which blocks a personal license or a license from somewhere else. Turn that on at the organization level and repository admins cannot turn it off. (October 8 review billing controls)

An automatic review is attributed to the pull request author by default. If the author is the Copilot cloud agent, usage goes first to the associated human co-author, then to the organization if that person cannot be billed. Reviews requested by other bots bill the organization directly. The person who clicked “request review” and the seat whose monthly allowance shrinks can now be different people.

Throughput: what to hand the agent, and what still needs a person

Throughput arguments usually mix three different products.

• Completions and Next Edit Suggestions stay included on every plan and do not consume AI Credits.
• Chat and multi-step coding-agent sessions consume Credits by token. A short question and a long autonomous coding run are no longer the same “request.”
• Code review consumes both AI Credits and Actions minutes. Minutes use the same per-minute rates as other Actions workflows. Self-hosted runners do not consume hosted minutes.

A March 18, 2026 changelog entry says the coding agent runs the project’s own tests and linters, then GitHub’s validation tools: CodeQL, the Advisory Database, secret scanning, and Copilot code review. Those checks are on by default, free, and do not require a GitHub Advanced Security license. Repository admins can turn individual tools off, for example when CodeQL is slow on one repo. (coding agent validation tools)

Keep the two charges apart. The validation pass the agent runs before it stops is described as free. A Copilot code review that a person requests on the pull request still uses the Credits and Actions minutes above. A free validation pass does not waive the review model.

The time that actually comes back is a shorter loop: the agent opens a change, tests fail, a review comment lands, and the suggestion goes back to the cloud agent. That is not the same as deleting human review. Security-sensitive and cross-service changes still belong on Balanced, with a person approving. Style, copy, and lockfile edits are usually a Lite job. Dependency manifests, logs, and SVG files are excluded from review anyway.

Frequency is the other trap. By default an automatic review runs once, the first time a pull request becomes Open. “Review new pushes” reviews every new commit. Eight round trips can become eight Balanced runs. Draft review is a separate switch. A user’s own setting and a ruleset do not override each other: if either is on, the pull request is reviewed; if both are on, Copilot still posts one review.

Teams already splitting cheap and expensive models can use the same habit described in how to use Claude Haiku 5.5 for agent subtasks. Bounded steps with obvious failures go on the cheaper path first. Cross-file reasoning and steps that change external state should not be automated from a product description alone.

AI coding cost: seat price held, the bill did not

On June 1, 2026, Copilot replaced premium requests with GitHub AI Credits metered on input, output, and cached tokens. Seat prices did not change: Pro is still $10 a month, Pro+ $39, Business $19 per user, Enterprise $39 per user. Included Credits match those prices: $10, $39, $19, and $39. Business and Enterprise included usage can be pooled, so one idle seat can cover a heavy user. When the allowance is gone, Copilot no longer falls back to a cheaper model. Budgets decide whether usage stops or continues at published rates. (usage-based billing announcement)

Existing Business and Enterprise customers received a higher transitional allowance in June, July, and August ($30 and $70 a month). That was a three-month promotional inclusion. In October, do not assume it is still there. Read the current allowance on the billing overview.

The arithmetic below is only a sense of scale, not a quote. Twenty Balanced reviews land somewhere around $5–$100 of AI Credits, before Actions minutes. The $10 included with Pro can disappear after a few Balanced reviews plus one long coding-agent session. Business at $19 is the same shape, except pooling lets light users subsidize heavy ones.

Members with no Copilot license, and members on Free, which does not include code review, have no review allowance of their own. If the organization enables both paid AI Credit usage and review for members without a license, that usage is billed to the organization as additional usage and is not charged to a personal budget. The option does not cover reviews inside an IDE.

For iOS and macOS teams this is also a runtime question. Copilot code review already includes Xcode. If the agentic step has to compile, sign, or boot a simulator, a GitHub-hosted Linux runner cannot do that job. A macOS self-hosted runner can. Moving those checks onto a dedicated Mac saves hosted Actions minutes and queue time, not Copilot Credits. Compare ZavCloud Mac cloud plans with the CI machines you already run before you treat a new runner as the cost plan.

Settings worth confirming before October 28

Each item below is already visible in organization or repository settings. Each one maps to a behavior GitHub has published, not to a guess about October 28.

• Effort: Confirm the repository is on explicit Lite or Balanced, not an untouched Default. Default has meant Balanced since September 28.
• Triggers: Decide whether every push and every draft is reviewed. Credits scale with review count.
• Who is billed: A member quota that fails closed, or the organization with a budget. Automatic reviews bill the author by default.
• Who may request a review: Whether external licenses are blocked. An organization-level block cannot be undone per repository.
• Approvals: Copilot approvals are still a preview and do not count as required approvals by default. Watch the dismissal behavior on a non-critical repo first.
• Validation tools: The coding agent’s CodeQL and related checks are free by default. Turn one off only when a repository is clearly slowed down. Do not disable all of them to avoid a license fee that is not being charged.
• Runners: Add a self-hosted Mac runner when the toolchain is macOS. It removes Actions minutes. Credits are separate.

After October 28, re-check these defaults against that day’s changelog. If seat prices, review ranges, or runner billing move, the budget sheet has to move with them.

Written on October 10, 2026. Dates, the Balanced default, Credit ranges, and the October 8 billing switches were checked against GitHub’s blog, changelog, and Copilot code review docs.

Questions teams keep asking

When is GitHub Universe 2026?

October 28–29, 2026, at Fort Mason Center in San Francisco, in person and online. As of October 10 the opening keynote had not happened. A session title is not a shipped feature.

What does one Copilot code review cost?

GitHub’s typical estimate is about $0.05–$1 of AI Credits for Lite and about $0.25–$5 for Balanced, excluding Actions minutes. It is a range, not a unit price. Multiply the range by how often you review before you decide whether it will eat the Credits included with a seat.

Can a self-hosted runner erase the review bill?

Not all of it. A self-hosted runner does not consume GitHub Actions minutes. The review model and coding-agent sessions still consume AI Credits. If hosted runners are disabled and no self-hosted runner is configured, agentic context gathering falls back to a more limited review.