GitHub Universe 2026: Copilot Fusion GA, the Copilot App, and the End of the Seat Fee

On October 28, 2026, GitHub turned its annual Universe conference in San Francisco into one of the most aggressive product resets in Copilot's history. The company's keynote post, written by Jay Parikh — Microsoft's EVP of CoreAI, whose organization now includes GitHub — framed it as a single vision: GitHub as "the most comprehensive software delivery system in the world," built for an era where developers orchestrate fleets of agents instead of writing every line themselves.

Buried under the vision talk are six concrete launches, one pricing change that will rearrange enterprise budgets, and a $5.59 benchmark result GitHub clearly hopes you remember. Here is what actually shipped, what it costs, and where the fine print lives.

What GitHub actually announced

The GitHub Copilot app: from issue to merge, in one surface

The headline product is a brand-new GitHub Copilot app, described by GitHub as designed from the ground up for agentic software delivery. It connects agents and tools across projects, and it deliberately goes beyond code: you can write specs, run design discussions, or analyze telemetry in what GitHub calls an "agent-first surface."

Two details from the launch coverage are worth noting. The app can generate and edit images from natural language in chat, and it has a computer-use feature — in public preview on macOS and Windows — that lets Copilot read app content, click controls, and navigate desktop apps, including ones with no API, CLI, or MCP integration. Copilot asks for approval before controlling an app, and organizations can disable the feature through managed settings.

Copilot Fusion: you stop picking models, it starts picking them

Fusion is GitHub's answer to model fatigue. Selecting Copilot Fusion in the model picker chooses a routing and orchestration system rather than a specific model. For each prompt, Fusion evaluates what the task needs across reasoning, code generation, debugging, and tool use, then decides how to execute across one or more models.

You pick an effort level — Standard for everyday coding, High for harder tasks, Max for the most demanding work — and Fusion runs one of three execution workflows: Single (one model solves the task), Cascade (an efficient model drafts, a quality check accepts or escalates to a stronger one), or Critique (one model drafts, a second reviews, the first revises). It combines the multi-model orchestration of GitHub's Project HydraFusion research preview — introduced September 4, 2026 — with the fine-tuned model selection behind Copilot's existing auto model selection. Developers who were on auto, HydraFusion, or HydraFusion Max during the preview move to the matching Fusion effort level automatically.

The model pool Fusion draws on includes Claude Opus 5.5, Claude Sonnet 5.5, GPT-6 Astra, and GPT-6 Sol, per reporting — a reminder that GitHub now routes across both Anthropic and OpenAI models inside one product. Administrators can restrict which models Fusion can use through enterprise and organization policies without disabling them for everyone.

Then there is the number GitHub is leading with. Citing the Artificial Analysis Coding Agent Index v2.0, GitHub says Fusion Max scored 60 — within two points of Claude Code running Opus 5.5 Max at 62 — at about half the average API cost per task, with about half as many tokens and roughly 40% less time per task.

The independent read is worth separating from the marketing. VentureBeat and AlphaSignal reported the index's own figures: Fusion Max at $5.59 per task versus $11.42 for Claude Code with Opus 5.5 Max (about 51% less), with 7.2 million tokens per task versus roughly 14.4 million. A competing orchestration system, Devin Fusion (Opus 5.5 Max plus GPT-6 Luna), also scored 62 at $8.00 per task. On the sub-benchmarks, Fusion actually led on Terminal-Bench 4.0 with 71, but trailed on FrontierCode (51 vs 56 for Claude Code). Important caveats: the 40% time saving is GitHub's claim — Artificial Analysis did not report per-task timing data — and Microsoft CEO Satya Nadella's promoted "up to 77%" savings figure refers to a separate, company-reported evaluation, not the index average. These are benchmark costs, not a quote for what your enterprise will pay.

Copilot Projects: one goal, many repos, parallel agents

Available natively in the Copilot app from October 28, Copilot Projects lets you run agent work across multiple repositories in fully managed, parallel sessions — one place to track progress, make decisions, and direct what happens next. You describe a goal, attach context (repositories, design mocks, even Slack threads), choose models and tools, and get results like pull requests back for review. It is in public preview, and GitHub's example of project automations — "create a promo video whenever a feature lands" — shows they want it to own repeatable work, not just one-off tasks.

Session continuity: your agent follows you

One of the most practical changes: you can start and follow the same cloud agent session across the GitHub app, VS Code, the CLI, and github.com. Start a session on your phone at standup, steer it from VS Code at your desk, check in from the web at lunch. The agent keeps working in the cloud. GitHub credits a multi-model harness for making this "powerful, fast, and cost-efficient."

Copilot Automations: workflows that rewrite themselves

Automations — AI-powered workflows inside a repository that run in the cloud, triggered manually, by events, or on a schedule — now learn from their outcomes and propose improvements for a human to review and approve. Issue triage and code review are the canonical examples. This is the learning-engine idea from Parikh's post in its most literal form: the output of completed work feeds the next run.

Copilot Code Review Max: the deep tier

GitHub added a third code-review tier. Lite and Balanced stay for everyday reviews; Code Review Max posted the highest F1 score on ReviewBench for high-severity issues, per GitHub. It is aimed at changes that deserve a deeper look before they ship — which, given that one write-up of the launch notes agents now start more than 40% of pull requests on GitHub, is most of them.

The money story: AI credits, not seats

The change that matters most to enterprise budgets is not a price cut — it is a pricing model change. Starting October 28, GitHub Enterprise customers can use all Copilot features — the app, code review, automations — without buying a separate Copilot subscription. You pay for the AI credits you use. It is available through GitHub Sales today, with self-serve coming soon. Existing Copilot Business and Enterprise customers can keep their current plans or move over when ready.

How credits work: 1 GitHub AI credit = $0.01. Every plan includes a monthly allowance. The plans page lists individual tiers at $10/month for Pro, $39/month for Pro+, and $100/month for Max, plus Free (limited chat and agent usage, 2,000 completions a month on entry-level models) and Business/Enterprise tiers with pooled or flex usage. Code completions and next-edit suggestions do not consume credits — they remain unlimited on paid plans. A quick question to a lightweight model costs a fraction of a credit; a long agent session on a frontier model across many files costs more.

Fusion's billing has its own rules: it is billed per model at each model's standard rate, and users on paid Copilot plans get a limited-time 30% discount on model costs when using it — reported to run through November 30, 2026. That discount is the difference between a demo and a commitment, so mark the date: after it expires, Fusion Max's cost structure is whatever the underlying frontier models charge, with no permanent GitHub discount.

On Business and Enterprise, admins set usage limits and decide whether additional paid usage is allowed — if not, Copilot pauses until the next cycle — with alerts at 75%, 90%, and 100% of any configured budget. And a privacy note the plans page buries in its FAQ: for individual (Free/Pro/Pro+) subscribers, GitHub may use Copilot interaction data — prompts, suggestions, and generated code snippets — to train and improve AI models. You can opt out in settings without losing features. Business and Enterprise data is governed by the Data Protection Agreement instead. For the full category picture, see our comparison of the 7 best AI coding agents in 2026, with prices, limits, and our pick.

The infrastructure behind it

GitHub is also spending on concrete: 4 million CPU cores, 160 PB of high-speed storage, and 50 Tb/s of network capacity added so far in 2026, plus a migration to a multi-region Azure deployment expected to complete in early 2027. More structurally, the company is rebuilding its Git architecture around three principles — simplify the primitives, decouple storage from compute, minimize coordination — to raise the throughput ceiling for an era when agents, not just developers, hammer the platform.

The scale numbers explain why. GitHub's Octoverse 2026 report, released October 27, puts India's developer community at 32 million. And in a neat bit of timing, former GitHub CEO Thomas Dhomke's new startup Entire launched a Git Forge rival earlier the same week — GitHub's "most comprehensive software delivery system" pitch reads partly as a response to competition coming from its own alumni.

Where it is strong, where the fine print lives

Strong: Fusion is the first mainstream attempt to make multi-model orchestration a checkbox instead of a research project — effort levels plus Single/Cascade/Critique workflows is a sensible interface for "just make it good and cheap." The independent index numbers are genuinely good: 60 vs 62 at half the benchmark cost, with a Terminal-Bench lead. The Enterprise pricing change is real money saved for companies that already pay both GitHub and Copilot seats. And Projects plus session continuity attack the actual pain of agent work — context scattered across repos, tools, and devices.

Fine print: the 30% Fusion discount is temporary and the end date (reported November 30) matters more than the discount. The 40% time saving and 77% savings claims are company-reported, not independently measured. ReviewBench's "highest F1" is GitHub's own citation — treat it like a vendor benchmark. Enterprise "no subscription" still means you pay for every credit, and there is no published estimate of what a heavy agent workflow costs under the new model. The model pool GitHub routes across includes competitors' models (Anthropic, OpenAI), which is great for users and a reminder of how much of GitHub's agent stack runs on models it does not train. And the whole package landed two days before most press coverage — expect the first wave of real-world Fusion cost complaints to define whether that 51% benchmark advantage survives production.

Frequently asked questions

What is Copilot Fusion, and does it cost extra?

Fusion is a multi-model routing and orchestration system you select in the Copilot model picker instead of a single model. It costs what the models it uses cost — billed per model at each model's standard rate — with a limited-time 30% discount for paid-plan users (reported through November 30, 2026). It is available on all Copilot plans in VS Code, the Copilot app, the CLI, and JetBrains IDEs.

Do I still need a Copilot subscription on GitHub Enterprise?

No — as of October 28, 2026, GitHub Enterprise customers can use all Copilot features without a separate subscription and pay only for the AI credits they use. It is available through GitHub Sales now, with self-serve coming soon. Existing Copilot Business and Enterprise customers can keep their plans or switch.

Can I just pick my own model instead of using Fusion?

Yes. Fusion is one option in the model picker; individual model selection remains. Admins can also restrict which models Fusion may use through enterprise and organization policies, or set a Fusion effort level as the default.

Does GitHub train models on my code?

For individual Free/Pro/Pro+ subscribers, GitHub may use Copilot interaction data — including prompts and generated code — to train and improve AI models, unless you opt out in your Copilot settings. Business and Enterprise usage is governed by the Data Protection Agreement, and there is a public-code duplication filter (65 lexemes or more) that can suppress suggestions matching public code.

Is the Copilot app's computer-use feature safe to turn on?

It is in public preview on macOS and Windows, asks for your approval before controlling any app, and organizations can disable it entirely through managed settings. Like any agent that clicks things on your desktop, it is worth testing on a spare profile before giving it broad access.

Key takeaways

• GitHub Universe 2026's centerpiece is Copilot Fusion, now generally available: pick Standard, High, or Max effort, and Copilot routes each task across models using Single, Cascade, or Critique workflows.
• Independently measured on the Artificial Analysis Coding Agent Index v2.0, Fusion Max scored 60 versus 62 for Claude Code with Opus 5.5 Max — at $5.59 per task versus $11.42, roughly half the cost with half the tokens.
• GitHub Enterprise customers no longer need a separate Copilot subscription; they pay only for AI credits used ($0.01 each), via GitHub Sales today with self-serve coming.
• The 30% limited-time discount on Fusion model costs, the 40% time-saving claim, and the "highest F1" ReviewBench score are all GitHub-reported — benchmark well, then budget from your own usage.
• Underneath the launches: 4 million CPU cores, 160 PB of storage, and a rebuilt Git architecture, all aimed at an era where agents start a large and growing share of pull requests.

As of October 30, 2026, GitHub's pitch is no longer "pick a model" but "pick an effort level and stop thinking about models." If Fusion's production bills track its benchmark numbers, that is a genuine step forward for agentic coding — and a pricing headache for every tool still charging by the seat. The next ninety days of real-world usage data will decide which story wins.