Copilot Fusion Is Here: GitHub's AI Model Router Delivers Near-Opus Coding at Half the Cost

What Happened: Fusion Went Generally Available at Universe

At GitHub Universe on October 28, 2026, GitHub made Copilot Fusion generally available across every Copilot plan — and quietly changed what it means to "pick a model" in an AI coding tool. Fusion isn't a model at all. It's a routing and orchestration engine: when you submit a task, it decides which AI model — or which combination of models — should handle each step, drawing on a pool that includes Anthropic's Claude Opus 5.5, OpenAI's GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna.

The release caps a two-month arc: GitHub shipped auto model selection earlier this year, launched the Project HydraFusion multi-model research preview on September 4, and has now merged both into a single product. Users who enabled auto, HydraFusion, or HydraFusion Max during the preview transition automatically to the corresponding Fusion effort levels.

The timing is no accident. GitHub's own Octoverse data, published alongside the launch, shows AI agents now start roughly 40% of pull requests on the platform. When agents write most of the code, the question stops being "which model is best" and becomes "who can route me to the right model per task, per step, at the lowest cost." Fusion is GitHub's answer — and it shipped as a default-feeling mode, not a settings toggle for power users.

How Fusion Works: An Orchestrator, Not a Model

Instead of choosing a model in the picker, you choose Fusion plus an effort level — Standard, High, or Max. Standard prioritizes versatile, cost-efficient models for everyday coding; High targets harder tasks with deeper reasoning; Max reserves the most capable models for the most demanding work.

Under the hood, Fusion evaluates what each task needs across reasoning, code generation, debugging, and tool use, then runs one of three execution patterns:

• Single — the task goes directly to the one model Fusion judges best for the job. No overhead when the answer is obvious.
• Cascade — an efficient model drafts a response, and a quality check either accepts it or escalates to a stronger model. Cheap first, expensive only when needed.
• Critique — one model drafts, a second model reviews, and the original revises. This is the closest pattern to how senior engineering teams actually work.

As GitHub's Mario Rodriguez put it: "It's not a model. It's an orchestration engine at runtime." The engineering insight is that different steps of one coding task have wildly different difficulty — a rename refactor doesn't need Opus-class reasoning, but a race-condition fix in a distributed system might. Fusion is the first mainstream product to price and route on that difference automatically.

The Numbers: Two Points Behind Opus 5.5 at Half the Price

Independent testing published October 28 by Artificial Analysis on its Coding Agent Index v2.0 gives Fusion its headline numbers:

That's within two points of the leading single-agent configuration at 51% less cost per task, using about half as many tokens, and finishing roughly 40% faster. GitHub's internal DeepSWE testing went further: Fusion High and Max beat comparable frontier-model configurations by up to 8.7 points on task completion while cutting estimated costs by as much as 77%. Microsoft CEO Satya Nadella's summary — frontier-level quality, 40% faster, up to 77% cheaper — matches the internal numbers even if the independent ones are more conservative.

Notably, Cognition's Devin scored its matching 62 with its own multi-model orchestration — a sign that routing, not any single model, is where the coding-agent competition is now being decided.

Code Review Max and the 40%-Agent Pull Request

Fusion wasn't the only coding announcement at Universe. GitHub also introduced Copilot code review Max, a new tier above the existing Lite and Balanced options, intended for changes that deserve a deeper look. GitHub says it posted the highest F1 score on ReviewBench for high-severity issues — the bugs that actually matter — and paired it with AI-assisted security scanning designed to catch vulnerabilities that traditional static analysis misses.

There's also a new GitHub Copilot app and Copilot Projects, rounding out a stack where agents draft, review, and ship more of the average repository every month. If agents start 40% of pull requests and AI reviews the rest, the human's job shifts toward specifying, approving, and owning outcomes — which makes the quality of the review tier a bigger deal than it first appears.

Pricing: Per-Model Billing With a 30% Discount

Fusion is available now in VS Code, the GitHub Copilot app, Copilot CLI, and JetBrains IDEs, with more clients planned. Billing is per underlying model at that model's standard rate — you pay for what Fusion actually calls. Paid Copilot subscribers get a limited-time 30% discount on model costs when using Fusion, separate from the task-cost savings measured in benchmarks. Enterprise admins keep full control: Fusion only uses models allowed by organizational policy.

What Model Routing Means for Your Coding Stack

If you're evaluating AI coding tools this quarter, Fusion changes the calculus in three ways:

1. The unit of comparison is shifting from model to harness. The same Opus 5.5 performs differently inside different tools. When Fusion hits 60 by mixing models, leaderboards of individual models tell you less than ever about your actual experience. Compare tools on your workload, not on model marketing.

2. Cost-per-task is the metric that matters. $5.59 versus $11.42 for near-identical output quality compounds fast on a team doing hundreds of agentic tasks a day. Ask vendors for cost-per-task numbers, not just benchmark scores.

3. Single-model loyalty is now expensive. Between Fusion, Devin's orchestration, and comparable routing in tools like Cursor, Windsurf, and Claude Code, the market is converging on the same conclusion: route across models instead of betting on one. The main risk is lock-in to one vendor's router — keep workflow definitions portable.

You can compare GitHub Copilot against Devin, Claude Code, Cursor, and the rest of the field on aitrove.ai, with pricing and alternatives for each.

Frequently Asked Questions