GitHub stops asking developers to pick a model: Copilot Fusion and Agentic Workflows go generally available at Universe

For years the ritual was the same: open the model picker, weigh speed against smarts against price, pick one, and hope you picked right. At GitHub Universe on October 28, GitHub announced it wants that ritual gone. Two of its biggest agent bets — Copilot Fusion and GitHub Agentic Workflows — reached general availability on the same day, and together they describe a different job for the developer: set the goal and the effort level, then review what the agents did.

Neither announcement is a new model. That is the point. GitHub, owned by Microsoft, is betting that the scarce resource in software teams is no longer raw model capability but orchestration — choosing the right model, running agents safely inside real repositories, and keeping the bill legible.

Fusion: pick an effort level, not a model#

Copilot Fusion folds together two earlier pieces of work — the HydraFusion research preview, which ran tasks across several models, and Copilot’s automatic model selection — into one setting. The developer chooses only how hard Copilot should try: Standard for everyday coding, High for harder problems and deeper reasoning, or Max for the most demanding work. Fusion then decides which models to use and how to combine them.

It has three coordination strategies. In Single, one model carries the task start to finish. In Cascade, one model plans and another executes, with a quality check that can escalate the work. In Critique, one model proposes an approach and a second reviews and challenges it before anything is written. The pool spans rival labs — Anthropic’s Claude Opus 5.5 and Sonnet 5.5 alongside OpenAI’s GPT-6 Astra and GPT-6 Sol — and Fusion runs in VS Code, the Copilot command-line tool, the Copilot app, and JetBrains editors.

On billing, there is no extra orchestration fee: teams pay for the underlying models Fusion invokes, with a 30% discount through November 30, 2026. Enterprise administrators get policy controls over which models Fusion may draw on, and can make it the default.

The benchmark numbers — and what they don’t prove#

GitHub’s headline claim comes from the Artificial Analysis Coding Agent Index, an independent benchmark of coding agents. Fusion Max scored 60, two points behind Claude Code running Opus 5.5 Max at 62 — while costing $5.59 per task against $11.42, using about half the tokens and finishing roughly 40% faster. On a separate internal test, GitHub says Fusion High and Max beat comparable frontier-model efforts on task completion by up to 8.7 points while cutting estimated costs by up to 77%.

Those figures deserve the usual caution. They are benchmark results under test conditions, reported by the vendor, and the index score is a composite — not a percentage of tasks completed. The costs are API estimates per benchmark task, not a promise about any team’s monthly bill. GitHub has also published this head-to-head comparison only for the Max tier. The real test is what Fusion costs and completes on ordinary repositories.

Agentic Workflows: repo chores as scheduled jobs#

The second announcement targets the maintenance work that eats engineering weeks. GitHub Agentic Workflows, an open-source project from GitHub Next, lets teams describe a repository job in plain-language Markdown — triage new issues, investigate a failing build, refresh documentation, write tests — and compiles it into an ordinary GitHub Actions workflow. The agent runs inside Actions, reusing the team’s existing runners, branch protections, and policy constraints. A command-line extension, gh aw, does the compiling; Copilot is the default engine, with Claude Code, OpenAI Codex, and Google Gemini CLI also supported.

Because these workflows act without a human in the loop, GitHub has wrapped them in safeguards: agents run with read-only permissions by default inside a sandboxed container behind what GitHub calls the Agent Workflow Firewall, outputs pass through a safe-outputs check, and a threat-detection job scans proposed changes before they land. The general-availability release adds Actions Runner Controller and Docker-in-Docker support, per-workflow model providers and model routing, and stronger isolation options.

GitHub’s showcase example is a single Accenture team, quoted in the announcement: one lead reportedly saved an estimated 29 hours a week, velocity more than doubled, AI now authors 91% of the team’s commits, and production errors fell from about 11,000 a day to under 1,000. One team’s story, told by the vendor — but it explains the push.

The platform play underneath#

Fusion and Agentic Workflows did not arrive alone. GitHub put Copilot Projects into public preview — a persistent workspace grouping context, repositories, agents, and outputs around a goal, including work spanning several repositories — and made its Copilot harness available through Microsoft Foundry, so other developers can build on the same orchestration layer. For enterprises, Copilot is now included for every licensed user, with billing shifting to the AI actually consumed rather than a separate per-seat licence.

The scale argument came from GitHub’s Octoverse data: by September, more than 40% of pull requests started with an agent’s first draft, up from 17% in March. GitHub says it has added four million CPU cores and 160 petabytes of storage this year to absorb that load. Agents are no longer a feature on top of GitHub; they are the traffic.

What to watch#

Three things will decide whether this lands. First, the bill: whether Fusion’s benchmark efficiency survives contact with real codebases, and whether teams can predict what an escalated Max task costs before it runs. Second, trust: whether the workflow safeguards hold up once agents are triaging issues and opening pull requests at scale, and how quickly teams catch the first bad autonomous change. Third, control: administrators now hold the model-policy levers, and the tension between org-wide defaults and individual developers’ preferences will only grow.

Rivals will not stand still: Anthropic’s Claude Code and OpenAI’s Codex each have their own routing and agent stories. What GitHub has that they do not is the repository itself. Universe 2026 was its bid to turn that home-field advantage into the default way software gets built.