GitHub made Copilot Fusion generally available on October 28, 2026, at its Universe conference in San Francisco, and in the same announcement tied Copilot to GitHub Enterprise so that every Enterprise seat carries it without a second subscription, according to GitHub’s keynote post by Jay Parikh. The two changes point in one direction: the model is no longer the product, and the meter is no longer the seat. A developer picks an effort level, a router picks the models, and the bill follows the AI credits consumed.
Video: “Work together, your way, with GitHub Copilot,” a GitHub Universe 2026 session published on the GitHub channel on YouTube. The figures in this article come from the written sources listed below.
What Fusion actually does
Selecting Fusion in a Copilot model picker does not select a model. It selects a routing system. According to Unite.AI, the system works at three effort levels: Standard for everyday coding, High for harder tasks, and Max for the most demanding work. Behind each level sit three execution patterns. A single model can handle the whole task. A cascade can have a cheaper model produce a draft, run a quality check, and escalate to a stronger model only if the check fails. A critique pattern has one model draft, another review, and the first revise.
The design traces to Project HydraFusion, a research preview GitHub released on September 4, 2026. The same Unite.AI report lists offline results GitHub published with that preview: on TerminalBench 2.1, a gain of 4.9 points in quality at 67 percent lower estimated cost than Claude Opus 5; on DeepSWE, a result within 1.5 points at 36 percent lower cost; and on CheckpointBench, within 0.1 points at 65 percent lower cost. Preview users of the auto, HydraFusion and HydraFusion Max settings move automatically to the matching Fusion level.
Administrators keep a hand on the wheel. They can restrict which models Fusion may choose, set policy per model, and decide whether Fusion is the default for their Copilot users, as The New Stack reports. Availability at launch covers all Copilot plans in VS Code, the GitHub Copilot app, the Copilot CLI and JetBrains IDEs, with more clients promised.
The benchmark claim and who made it
The headline number comes from GitHub itself. Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code running Opus 5.5 Max. GitHub says it used about half the average API cost per task, about half the tokens and about 40 percent less time. The New Stack notes that these claims come from GitHub and one customer engineer, with no independent verification or full test methodology attached, and that “half the cost” describes API cost per task, not what a developer pays for a Copilot plan.
Executives have not settled on a single cost figure, either. At Universe, Satya Nadella said Fusion delivers frontier-level quality 40 percent faster and at up to 77 percent lower cost, according to Digg. Parikh’s figure was half the cost. The two numbers measure different things, and the Next Web report tells readers to treat them as separate claims.
One customer voice supports the routing idea from the ground. Hari Pokhrel, a principal software engineer at AT&T, credited Fusion Standard with better token efficiency and said developers “forget to switch back for everyday activities” after choosing a powerful model for one hard job. That habit is the waste a router targets, and it is a behavioral claim rather than a benchmark.
The routing logic also fits what developers were already doing. In the Octoverse data, 61 percent of Copilot users who chose their own model in September 2026 used more than one, up from 50 percent four months earlier. Paid plans drew on 47 distinct models that month, against 29 in May, and no single model carried more than 25 percent of requests. By September, 63 percent of paying users had used automatic model selection at least once. A tool that picks among dozens of models is answering a problem that users had created for themselves.
From seats to credits
Starting October 28, GitHub Enterprise customers can use every Copilot feature, including the Copilot app, code review and automations, without a separate subscription, and they pay for the AI credits they use. GitHub says access runs through its sales team now, with self-serve “coming soon,” and that existing Copilot Business and Enterprise customers can keep their plans or switch. No credit rates were published, and self-serve purchasing is not yet available.
Fusion is billed per model at each model’s standard rate, and paid Copilot users receive a limited-time discount of 30 percent on model costs when using it, which ITWire reports runs through the end of November. A discount with an end date makes the first month of use a poor guide to the steady-state bill. Teams that want a clear read should wait for the discount to lapse before judging cost.
For a sense of how token prices behave in practice, x9m8’s earlier calculation of Claude Haiku 5.5 pricing shows how a per-token rate turns into a per-task cost once a workload is specified. Credit-based billing adds one more layer between those two numbers.
Projects, automations and review in one app
The GitHub Copilot app is described as an agent-first surface that connects agents and tools across projects and supports writing specs, design discussions and telemetry analysis. Copilot projects, available in the app from October 28 as a public preview, runs agent work across multiple repositories in managed parallel sessions with one place to track progress. A session started in the app can be followed in VS Code, the CLI or on the web.
Copilot automations are cloud-run workflows in a repository that can start manually, on an event or on a schedule, for jobs such as issue triage. Automations can now learn from their results and propose changes, but the proposals wait for a developer to approve them. A new Copilot code review Max tier, which GitHub says posted the highest F1 score on ReviewBench for high-severity issues, sits above the Lite and Balanced tiers. A computer-use feature is in public preview on macOS and Windows; it requires user approval and managed settings can turn it off, per Unite.AI.
The keynote demo made the pitch concrete. ITWire describes a fictional coffee loyalty app for a chain of 35 cafes, built by Cassidy Williams inside the Copilot app. The assistant broke the idea into tasks, pulled in Slack and Teams discussions and designs, drove a browser, ran parallel agents and produced promotional material. GitHub COO Kyle Daigle argued in the same coverage that producing code is now comparatively easy, and that the harder enterprise problem is getting trustworthy, well-governed software into production.
Permission design matters more as the agents gain reach. x9m8’s report on the Anthropic agent containment findings covers what happens when an agent takes actions nobody asked for, which is the case that approval gates and managed settings exist to catch.
What the Octoverse numbers show
The Octoverse 2026 report covers October 1, 2025 through September 30, 2026. GitHub counted 17.9 billion pushes, against 6.7 billion in 2025 and 5.2 billion in 2024. Commits reached 34.3 billion, up from 12.9 billion. Merged pull requests came to 1.3 billion, or about 3.5 million per day, and GitHub Actions ran 45.4 billion jobs. The platform has more than 225 million developers, with 38.3 million in the United States and 32 million in India.
Agents drive much of the growth. More than 40 percent of pull requests in September 2026 started with an agent, up from 17 percent in March. About 110 million agent-started pull requests landed in September, roughly six times March’s volume. The report adds that agent-started pull requests changed multiple files about 87 percent of the time, against 79 percent for those without detected agent involvement, and that the share of agent-involved pull requests merged within two weeks rose from 83 percent in February to 90 percent in September.
The measurement carries a caveat that GitHub states itself. An agent-started pull request is one whose first draft came from a coding agent the platform can detect, and the report calls the signal “a floor, since not all agent use may be detected.” Repositories without an agent configuration file may still use AI. Among repositories created in 2026, those with such a file logged eight times the commits of those without one.
Language data points the same way. TypeScript led contributor rankings, followed by Python and JavaScript, and Rust doubled in a year, the fastest rate among top languages. Detected AI involvement in merged pull requests ran from 44.1 percent in TypeScript to 21.2 percent in Java, the lowest of the languages shown, and above one in five in every language listed. The Copilot figures that point toward multi-model use come from the same report, which says its counts describe observed use rather than model quality.
The load that agents put on Git
Machine-speed writing has a cost on the server side. GitHub told the Universe audience it had added 4 million CPU cores, 160 petabytes of high-speed storage and 50 terabits per second of network capacity during 2026, and that a multi-region Azure migration is expected to finish early next year. Japanese trade coverage by Cloud Watch reported that GitHub showed a chart of explosively rising load, tied it to coding agents committing code at speed, and said the load has caused outages. GitHub’s own related post on its September availability notes two incidents that degraded performance.
Coverage in CodeZine put the volume at more than 7 billion commits a month and said GitHub presented a new Git architecture built for heavy write load from agents. The keynote post states three principles for it: simplify the primitives, decouple storage from compute, and minimize coordination. Those details come from the visible portion of a member-gated article, so the architecture itself remains thinly documented in public.
Code that was once written by people at human pace is now generated, branched and merged by software around the clock. Whether the infrastructure keeps pace is a question the next availability report will answer better than any keynote.
Questions the launch leaves open
Several items remain unverified. The benchmark comparison rests on one index and on GitHub’s own run. Pricing is credit-based with no published rates, and self-serve purchasing is not yet available. The Copilot projects feature is a public preview, and the Azure migration is unfinished. x9m8’s coverage of the open-weight Kimi K3.1 Pro release is one example of the cheaper models a router could draw on, though GitHub has not said which models Fusion uses beyond what admins permit.
Independent testing will settle the cost and quality question. Until it arrives, the clearest verified facts are the structural ones: a router now stands between the developer and the model, and the unit of billing has moved from the seat to the credit.
Sources
- GitHub Blog, GitHub Universe 2026: Evolving every layer of the stack, October 28, 2026
- GitHub Blog, Octoverse: Commits, pushes reach new highs in 2026 as developers build with agents
- The Next Web, GitHub launches Copilot Fusion as agents start 40% of pull requests, October 28, 2026
- ITWire, GitHub announces new Copilot Fusion and Copilot Projects
- Unite.AI, GitHub Makes Copilot Fusion Generally Available at Universe 2026
- The New Stack, GitHub debuts Copilot Fusion, October 28, 2026
- Digg, GitHub highlights Copilot Fusion and multi-agent Projects at Universe
- Cloud Watch (Impress), GitHub Universe 2026 opens, Copilot Fusion GA (Japanese)
- CodeZine, GitHub Universe 2026 opening keynote report (Japanese, partly member-gated)
