GitHub Copilot Fusion Goes GA With Agents Starting 40% of PRs

A router, not a model

GitHub's main announcement at Universe 2026 was not a new model. It was software that chooses between models. Copilot Fusion became generally available on October 28. Developers pick an effort level, and Copilot decides which models handle each part of a coding task, so nobody has to choose a model for every request.84 The coverage consistently describes it as a router and coordinator rather than a standalone model. Fusion checks whether a request needs reasoning, code generation, debugging or tool use, then combines the output into one final answer.16

GitHub announced Fusion alongside its Octoverse 2026 data. The report says agents started more than 40% of pull requests on the platform in September, up from 17% in March. That works out to about 110 million agent-started pull requests in one month.2221 Put together, the two announcements show where the industry is heading. The hard parts of AI coding are moving away from writing code and toward deciding which models to use, reviewing what agents produce, and keeping the tools themselves trustworthy.

How Fusion works

There are three effort levels you can select. Standard handles everyday work, High covers larger and well-scoped jobs such as multi-file changes and difficult bug fixes, and Max targets the deepest reasoning. A fourth mode, Fusion Fast, runs speed-sensitive inline chat in VS Code and can't be selected manually.26 For each task, Fusion uses one of three patterns. Single sends the work to one model. Cascade lets a cheaper model draft an answer, then has a judge either accept it or pass the task to a stronger model. Critique has one model draft while others review, and the original model then revises.65

Fusion has been through several versions this year. A community timeline traces it from per-request Auto selection in January, to intent-scored routing in May, to per-turn orchestration under the Project HydraFusion name.7 That research preview began in Copilot CLI on September 4 and reached general availability roughly four weeks after its expansion to VS Code.74 By one account, HydraFusion routed almost 60,000 sessions before launch.3

The reasoning behind Fusion is that developers already spread their work across many models. Paid Copilot users used 47 different models in September, and no single model handled more than a quarter of their requests.3 GitHub also says 61% of users who picked models themselves used more than one.24 If developers are already switching between models, GitHub is betting it can switch for them more cheaply.

The benchmark claim and the Claude Code comparison

GitHub compares Fusion directly with Anthropic's Claude Code. On the Artificial Analysis Coding Agent Index, Fusion Max scored 60, against 62 for Claude Code running Opus 5.5 Max. GitHub says Fusion Max did this at about half the cost per task, with about half the tokens and roughly 40% less time.856 The coverage doesn't fully agree on the size of the savings. Microsoft CEO Satya Nadella described the gains as 40% faster and up to 77% cheaper, which goes beyond GitHub's own "half the cost" claim.8 Earlier figures from the HydraFusion preview cited a 67% lower estimated cost than Claude Opus 5 on a different benchmark. Analysts point out that those were vendor-reported offline evaluations, not independent tests of the released product.10

The more skeptical outlets are right to be careful. The headline number comes from GitHub's own announcement, and no independent reproduction is available yet.455 GitHub also hasn't published which models are in Fusion's pool.55 Fusion probably will save money on routine work, because most coding requests don't need a frontier model. Whether it matches Claude Code on hard problems hasn't been shown yet.

Pricing could also catch users out. Fusion has no separate fee and bills each model it uses at that model's standard rate, so High and Max can use more AI credits because several models may work on one task.2 Paid plans get 30% off Fusion model costs until November 30.4 That discount makes the early weeks look cheaper than ongoing use may turn out to be.

Agents write the drafts, people review them

The Octoverse numbers are large. In the 12 months to September, developers pushed code 17.9 billion times, made 34.3 billion commits and merged 1.3 billion pull requests.24 Agent-started pull requests touched multiple files 87% of the time, compared with 79% for other pull requests. Their two-week merge rate rose from 83% in February to 90% in September, while pull requests without a detected agent stayed around 86%.2225

The coverage agrees on how to read these numbers carefully. GitHub's definition counts any pull request whose first draft came from an agent it can detect, even if a human rewrote it afterwards. GitHub calls the figure a floor, because some agent use goes undetected.26 A merge rate shows how fast code gets accepted, not whether it is correct. The report doesn't measure defects, reverts or incidents in agent-written code.2928 Some of the volume may also be mechanical work. Dependabot alone merged more than 30 million pull requests over the year.28 One security vendor summed up the situation bluntly: code is now being written about three times faster than anyone can review it.30

That explains why AI code review was such a large part of the launch. GitHub added Max as a third effort level for Copilot code review, in public preview. It runs several independent review agents on models from different providers, then uses an aggregation agent to merge their findings into one review.52 Copilot approvals also became generally available. Once an admin enables them, Copilot's approval counts toward a repository's required reviews, can be limited to specific file paths, and is dismissed when new commits arrive.51

Reports differ on what Max costs. InfoWorld and a pricing analysis both cite GitHub figures of roughly $0.25 per review for Lite, $1.50 for Balanced and $10 for Max.5758 Another review of the documentation found no published price for Max, only that it costs more than Balanced.51 GitHub says Max posted the highest F1 score on ReviewBench for high-severity issues, but it hasn't released the scores behind that claim.51 Letting an AI approval count toward a merge is probably the bigger change than Max itself, because it hands part of the merge decision to a model.51

Claude Code and Cursor are building similar things

The competition is heading the same way. Anthropic made its own Code Review product generally available on October 15.37 Claude Code has shipped point releases at a fast pace. Version 2.1.302 made forked subagents the default across the SDK, and version 2.1.296 let teams run every subagent in a workflow on one chosen model.3236 That last feature is a manual counterpart to Fusion's automatic routing. Anthropic gives developers explicit control over models, while GitHub asks them to let Copilot decide. Other releases are making Claude Code more autonomous, with auto mode now the default for interactive sessions and new extension points called mods.3437

Cursor is turning into a platform for running agents as much as an editor. It released its Agent SDK on October 20, the same framework behind its Bugbot and Security Review agents, so teams can define agents in code and test them with evals.1216 On October 23 it added Computer Use to the local editor, which lets agents click through desktop apps such as the iOS Simulator.19 Reports on Cursor's cloud environments disagree: Cursor's release notes say they support up to 300 repositories, while another report says 500.1215 Cursor also says autonomous cloud agents now create more than 35% of the pull requests its own engineering team merges.15 That internal figure is close to GitHub's platform-wide number.

A supply-chain incident the same week

The week also showed how exposed these tools are. On October 29, a tampered build of the Claude Code extension, version 2.1.313, sat in Cursor's VS Code extension marketplace for about 11 hours. Cursor says it reached at most 5,000 users and was not built by Anthropic.4341 The build tried to add an attacker's SSH key, open a reverse tunnel and stay on the machine after restarts. Cursor says a key step would have failed on most machines because it needs admin rights.4943 Reports don't fully agree on the details. One outlet said 5,000 users were "already impacted", though Cursor's own statement gives that number as an upper limit.5042 One analysis placed the bad build on Open VSX too, while researchers cited elsewhere found the Open VSX, Microsoft marketplace and npm versions clean.4649

What this means

Agents now write a large share of first drafts, and the vendors know it. GitHub's answer is to manage cost through routing and to tackle the review backlog with more AI review, including AI approvals that count toward merges. That approach is coherent, but it means checking agent code with more agents. The tampered extension is a reminder that the tools at the center of this workflow are now targets themselves. Teams should treat the 40% figure as a reason to invest in review capacity and to vet the extensions and tools agents rely on, rather than as evidence that agent code is ready to ship unchecked.