What is GitHub Copilot Fusion? Pick an effort level, not a model — 2 points behind Opus 5.5 at half the cost

Copilot Fusion hands the job of choosing an AI model from the developer to Copilot itself. GitHub made it generally available on October 28, 2026 (US time) at its annual GitHub Universe conference. Instead of picking Claude or GPT from a model list, users pick Fusion and an effort level — Standard, High or Max. Copilot then decides, for each request, which models to use, how many and in what order.

GitHub led with two numbers. Its top tier, Fusion Max, scored 60 on an independent coding-agent benchmark, within two points of Claude Code running Claude Opus 5.5 Max — at about half the cost per task and about 40% less time. GitHub's annual Octoverse report, released the same day, explains why such a feature became necessary: agents started more than 40% of pull requests (PRs) on GitHub in September.

1. What Copilot Fusion is — a model picker turned into a router

Until now Copilot users chose a model from a drop-down. The list has grown long: Octoverse 2026 says the number of distinct models chosen by paying users rose from 29 in May to 47 in September. Fusion automates that choice per task. Aggregators such as OpenRouter already let developers call many vendors' models through one key; Fusion goes further by combining several models within a single task.

Preview users of auto selection, HydraFusion or HydraFusion Max are moved automatically to the matching Fusion effort level. Octoverse says HydraFusion handled about 60,000 sessions during its preview.

2. Three effort levels, three execution patterns

Fusion has two axes: the effort level the user sets and the execution pattern Fusion picks.

Cascade is where the savings come from. Most requests are handled well by a cheaper model's draft, and only those that fail the check go to an expensive one, lowering the average cost compared with sending everything to a top model. The same model can also behave differently depending on the harness around it.

Two caveats: Fusion has no context window of its own, so each step is limited by the model running it. And if Fusion discards a draft, file edits that draft already made stay in the workspace until the user reviews them (per CellCog).

3. Performance and cost — GitHub's comparison

Slightly lower score, half the cost. The index tests whether an agent can finish real repository tasks end to end. Scores for the same model can vary widely between benchmarks, and these comparisons use API prices, not subscription fees.

Code review got a new tier too. Copilot code review Max uses several high-reasoning models for the deepest review, and GitHub says it posted the highest F1 score on ReviewBench for high-severity issues. It is in public preview.

These are GitHub Docs estimates and rise with PR size. Reviews on private repositories also consume GitHub Actions minutes.

4. How you pay — AI credits

When GitHub announced the move to usage-based billing in April, it said a quick chat question and a multi-hour autonomous coding session costing the same was no longer sustainable. With agents consuming millions of tokens per task, per-request pricing stopped making sense. Fusion fits that model as a way to get similar results for fewer credits. High and Max can use more credits per task than Standard.

5. Octoverse 2026 — why a router became necessary

Octoverse is GitHub's annual report on developer activity. This year's period runs from October 1, 2025 to September 30, 2026; GitHub recalculated 2024 and 2025 on the same definitions.

Pushes in 2026 exceeded 2024 and 2025 combined. Simon Willison, writing from the event, noted monthly figures of 7.38 billion commits and 220 million merged PRs. Growth was concentrated in the second half: GitHub marks June as the inflection point, and from June about 55% of all commits came from brand-new repositories.

This table is Fusion's rationale. Developers were already switching models by task, and no model held more than a quarter of requests. GitHub automated the switch. It notes the agent figures are a floor, since not all agent use is detected, and that 40% refers to PRs an agent started — not the share of code written by AI.

TypeScript remained the most used language, followed by Python and JavaScript. Rust grew fastest among the top languages, up about 162% in a year.

6. Other Universe announcements

The infrastructure build-out is a response to the surge in agent-generated commits and pushes. Jay Parikh, Microsoft's executive vice president for CoreAI, who oversees GitHub, said nearly all software written so far will be a small fraction of what is written in the next five years.

7. Frequently asked questions

Is Fusion always cheaper? No. With many easy tasks, cascades cut cost; but Max with frequent critique loops can cost more than a single top model. The half-cost figure comes from one benchmark comparing Max with Claude Code.

Will Fusion call a model my company has blocked? No. It only uses models allowed by organization and enterprise policies, and admins can disable a model for Fusion alone.

Can free users use it? GitHub says Fusion is available on all Copilot plans. The 30% discount applies only to paid plans.

What changes for teams outside the US? Billing is in dollar-denominated AI credits, so setting budget limits matters for teams. When a budget runs out, features such as code review are blocked.

8. What remains unknown

• Benchmark detail: Scores and costs rely on GitHub and coverage citing it; Claude Code's score is reported as 61.7 or 62.
• Discount end date: GitHub says only limited-time; November 30 comes from secondary summaries.
• Model list: The changelog does not list which vendors' models Fusion uses, and the pool changes over time.
• Enterprise credit rates: Specific rates for the credits-only Enterprise option have not been published.