GitHub Universe 2026: the Copilot app becomes the agent front door, Fusion orchestration ships GA

GitHub opened its Universe 2026 conference at Fort Mason Center in San Francisco on October 28 by recasting itself as an “end-to-end software delivery system”: a Copilot app meant to be the single front door for agents, the general-availability release of Copilot Fusion multi-model orchestration, cross-repo Copilot projects, and an all-Copilot bundle for Enterprise customers with no separate subscription. The hardest number of the day belongs to Fusion: per GitHub’s changelog, Fusion Max scored 60 on Artificial Analysis’s Coding Agent Index — within two points of Claude Code with Opus 5.5 Max — while costing about half as much per task in average API spend, using about half the tokens, and finishing roughly 40% faster.

Key points

• Fusion ships three orchestration modes and three effort tiers. Single hands a task to one model; Cascade drafts with an efficient model and escalates to a stronger one if a quality check fails; Critique has one model draft, a second review, and the first revise. Effort runs Standard, High, Max. You are billed each model’s standard rate for whatever Fusion picks, and paid-plan users get a limited-time 30% discount on model costs.
• The Copilot app becomes the agent front door. Copilot projects, available “starting today,” run agent work across multiple repos in managed parallel sessions with one place to track progress; the same cloud session continues across the GitHub app, VS Code, CLI, and the web. Code review gets a Max tier that GitHub says posts the highest F1 on ReviewBench for high-severity issues, with Lite and Balanced staying on.
• Enterprise billing changes shape. GitHub Enterprise customers no longer buy a separate Copilot subscription — every Copilot feature is bundled, billed only through AI credits as work is consumed. Activation runs through GitHub Sales today; self-serve is “coming soon,” and existing Business/Enterprise subscribers can keep their plans or migrate.

Background

The narrative came from the official blog signed by Microsoft’s EVP of CoreAI: GitHub now calls itself “the most comprehensive, end-to-end software delivery system in the world,” and swaps lines of code for “completed work” as the measure of output. Behind the framing sits an infrastructure bill — GitHub says it added 4 million CPU cores, 160 PB of high-speed storage, and 50 Tb/s of network capacity this year, with a multi-region Azure migration expected to finish early next year. Fusion is not a cold start either: the changelog describes it as the merger of the HydraFusion research preview’s orchestration with Copilot’s automatic model selection, and preview participants who had “auto,” “HydraFusion,” or “HydraFusion Max” selected were auto-transitioned to the corresponding Fusion tiers.

The facts

The parameters below come from the changelog and the official blog post, both dated October 28.

Item	Official claim
Benchmark	60 on the Artificial Analysis Coding Agent Index, within 2 points of Claude Code with Opus 5.5 Max
Cost and time	About half the average per-task API cost, about half the tokens, ~40% less time (official)
Orchestration	Single (one model); Cascade (draft plus quality-gated escalation); Critique (draft, review, revise)
Effort tiers	Standard for everyday coding, High for deeper reasoning, Max for the most capable models
Billing	Each model at its standard rate; limited-time 30% model-cost discount on paid plans
Availability	All Copilot plans; VS Code, the Copilot app, Copilot CLI, JetBrains IDEs — more clients pending
Enterprise bundle	All Copilot features without a subscription, billed via AI credits; via Sales now, self-serve to come

Trying Fusion

Existing Copilot subscribers can switch on it right now: open the model picker in VS Code, the Copilot app, Copilot CLI, or a JetBrains IDE, pick Copilot Fusion, and choose an effort level. Enterprises have one more gate — Fusion only draws from models enabled in the org’s model policy, admins can disable a model for Fusion alone, and any tier can be set as the default through managed settings.

Reading the benchmark

The benchmark numbers deserve to be read as vendor-reported claims. First, the claim is “within two points”: Fusion Max did not beat Claude Code with Opus 5.5 Max, and the comparison is against the Claude Code harness-plus-model pairing, not the raw model. Second, the preview and GA results are not one curve. As Unite.AI pieced together, the September HydraFusion preview claimed a 4.9-point quality gain at 67% lower cost than Claude Opus 5 on TerminalBench 2.1, near-parity at 65% lower cost on CheckpointBench — a GitHub internal benchmark built from real Copilot sessions, which outside parties cannot replicate. Different baselines, different opponents, different metrics: they do not compose into a “cost keeps falling” trend line.

What others say

The Verge condensed the Copilot app update to three things: assign Copilot a task across multiple repos, pick work up on another device, and build reusable AI automations inside the app. Unite.AI restated every headline number and quoted an AT&T principal engineer saying Fusion Standard “improved token efficiency across AT&T” and cut unnecessary spend. The same report notes the app carries a computer-use preview that can click, type, and navigate macOS and Windows applications that expose no API or MCP — per its account, Copilot asks for approval before controlling an app, and organizations can disable the feature through managed settings. And the official blog closed with a bigger claim: all software written to date will be less than 1% of what gets written in the next five years. One caveat of our own: the GA Fusion benchmark figures rest on GitHub’s word alone (see the previous section); no independent replication exists yet.

Our take

For developers, the practical change is small but real: the model picker’s question shifts from “which model” to “which budget,” and GitHub now does the routing for you. The trade is transparency — Fusion bills whatever model it picks at that model’s standard rate, and once the 30% discount lapses, heavy users’ bills track usage directly. Keep the usual skepticism toward launch-day benchmarks: parity on an internal benchmark and “close, not ahead” on a third-party index is standard release-season choreography. The thing actually worth watching is the Enterprise bundle: GitHub stops selling Copilot by the seat and starts metering how much work agents do through AI credits — that is a new revenue line, not a pricing tweak. Our earlier coverage of OpenAI Codex moving from cloud to local is the other side of the same coin: OpenAI pushes work onto the developer’s machine, GitHub pulls it into its own cloud sessions. And the benchmark’s foil, Claude Opus 5.5, is not standing still either.