Skip to content
FrontierDot

GitHub Copilot Fusion Timeline: What the Oct 28 GA Pages Actually Verify

Live timeline · last updated · 1 entry in the update log

Fusion is not a model you pick. It is a system that picks models for you. The useful controls are the effort level and which models your admin left enabled.

Timeline

On October 28, 2026 GitHub made Copilot Fusion generally available: a runtime router and orchestrator with Standard, High and Max effort levels, Single/Cascade/Critique execution patterns, billing at each underlying model’s rate plus a temporary 30% discount on paid Copilot plans through November 30, 2026, and experimental local-model routing that still decides in the cloud.

Read on 2026-10-30 (UTC+8): GitHub Changelog “Route and orchestrate models with GitHub Copilot Fusion” (Oct 28, 2026), docs.github.com “About GitHub Copilot Fusion”, the Oct 28 “GitHub Copilot Fusion now supports selection of local models” changelog, and @github status 2115480207469560109. Artificial Analysis’s Oct 28 article is cited only as AA-reported for Coding Agent Index figures. We did not run Copilot Fusion ourselves.

Timeline

  1. GitHub publishes the Copilot Fusion generally available changelog. Fusion combines Project HydraFusion’s multi-model orchestration with Copilot auto model selection. Users pick Fusion in the model picker, then choose an effort level: Standard, High or Max. Execution patterns named on the page are Single, Cascade and Critique. Fusion is available on all Copilot plans in Visual Studio Code, the GitHub Copilot app, GitHub Copilot CLI and JetBrains IDEs, with more clients “coming soon.” Billing is for each model Fusion uses at that model’s standard rate, with a limited-time 30% discount on model costs for paid Copilot plans.

    GitHub Changelog: Route and orchestrate models with GitHub Copilot Fusion

  2. GitHub’s Copilot Fusion docs page describes Fusion as an intelligent runtime model router and orchestrator. It adds Fusion Fast as a non-user-selectable effort level used only in Visual Studio Code inline chat. Cascade is framed as draft-then-judge-with-possible-escalation; Critique drafts with parallel reviewers whose feedback is aggregated before the original model revises. The same page states that through November 30, 2026, users on paid Copilot plans receive a 30% discount on model costs when using Copilot Fusion, and that Fusion has no context window of its own — each step uses the context limits of the model running it.

    GitHub Docs: About GitHub Copilot Fusion

  3. A second changelog ships intelligent routing for local models into experimental preview. On supported Windows devices, MAI Code 1.1 Flash can be downloaded through Windows ML; Copilot CLI can also use manually configured MAI Code 1.1 Flash or Qwen3.8 Flash Next on Windows, macOS and Linux. Local inference does not make the session offline: Copilot Fusion still makes routing decisions in the cloud. Local models are more likely to be used on Fusion Standard when eligible.

    GitHub Changelog: Copilot Fusion now supports selection of local models

  4. @github posts that in the Artificial Analysis Coding Agent Index, Copilot Fusion Max scored within two points of Claude Code with Opus 5.5 Max at about half the API cost per task, and points readers to a 30% off offer through November 30.

    @github on X: Copilot Fusion Max and 30% off (Oct 28)

  5. Artificial Analysis publishes its Coding Agent Index write-up on Copilot Fusion. AA reports Fusion Max at 60 versus Claude Code with Opus 5.5 Max at 62, about $5.59 per task versus about $11.42, roughly half the tokens, a Terminal-Bench 4.0 score of 71, FrontierCode at 51 and SWE-Atlas-QnA-AA at 57. These figures are AA-reported, not GitHub’s own table.

    Artificial Analysis: Copilot Fusion Coding Agent Index write-up (Oct 28, 2026)

What Fusion is, in GitHub’s own words

The GA changelog opens with a plain claim: instead of picking a model for each task, you choose Copilot Fusion and let Copilot route and orchestrate which models it uses. The docs page sharpens that into a product definition — “an intelligent runtime model router and orchestrator” that chooses models and execution patterns, then returns a single response.

That framing is the part most launch coverage softens. VentureBeat and Artificial Analysis lead with the Coding Agent Index scoreboard. Creators lead with “half the cost of Opus.” Both angles are real secondary stories. The primary product change is that the model picker entry is a system, not a weights file. For each prompt, Fusion picks an execution pattern, a reasoning effort for that step, and the models that run it. Different prompts in the same session can take different patterns.

GitHub says Fusion also watches real-time system health and availability, and that routing prefers natural cache boundaries so conversation stays on one model when that still makes sense. Intermediate drafts can be revised or discarded; the UI is meant to show the final response, with a hover that lists models used under the effort-level name.

Effort levels and the three execution patterns

User-facing effort levels on the changelog and docs are Fusion Standard, Fusion High and Fusion Max. Standard prioritizes versatile models for everyday coding. High prioritizes capable models for harder work. Max prioritizes the most capable models for the deepest reasoning. Docs add Fusion Fast, which is not selectable by users and exists only for Visual Studio Code inline chat.

Execution patterns matter more than the marketing nickname. Single sends the task to one model. Cascade lets an efficient model draft, then a judge accept or escalate to a stronger model. Critique has one model draft while other models critique in parallel; feedback is aggregated and the original model revises. Cascade and Critique take longer because they add review calls. Fusion only adds those calls when the task warrants them, according to the docs.

Administrators keep a separate lever. Fusion only uses models enabled in enterprise and organization policies. Admins can disable a model for Fusion without disabling it for everyone, and can set Fusion at any effort level as the default under Enterprise managed settings. If none of Fusion’s models are available to you, Fusion does not appear in the picker.

Billing, the November 30 discount, and what the AA score actually measures

Billing is not a flat Fusion fee. Both the changelog and the docs say you are billed for each model Fusion uses, at that model’s standard rate. High and Max can spend more AI credits than Standard because they may call more than one model. Through November 30, 2026, paid Copilot plans get a 30% discount on those model costs when using Fusion — the end date is printed on the docs page; the changelog calls it a limited-time discount without repeating the calendar day. @github’s Oct 28 post also says “Get 30% off through Nov. 30.”

GitHub’s changelog cites the Artificial Analysis Coding Agent Index and says Copilot Fusion Max scored 60, within two points of Claude Code with Opus 5.5 Max, at about half the average API cost per task, about half as many tokens and about 40% less time. The Artificial Analysis article is the place those index rows are spelled out as AA-reported numbers: about $5.59 per task versus about $11.42, Terminal-Bench 4.0 at 71, FrontierCode at 51, SWE-Atlas-QnA-AA at 57, and roughly 7.2M tokens per task. AA also names models Fusion can choose among, including Claude Opus 5.5, Claude Sonnet 5.5, GPT-6 Astra and GPT-6 Sol — that pool can change, and Fusion still respects your plan and admin policies.

Read the 60-versus-62 headline as a composite. AA weights FrontierCode, Terminal-Bench and SWE-Atlas equally. Fusion’s lead on Terminal-Bench does not erase the trailing rows on the other two. GitHub’s “about 40% less time” claim is GitHub-reported in the changelog; the AA article we read does not reprint that latency figure as its own measurement.

Local models: experimental, and still not offline

The local-models changelog puts intelligent local routing in public preview. Windows devices that meet the requirements can download MAI Code 1.1 Flash through Windows ML inside Copilot CLI, the Copilot app or VS Code. Copilot CLI can also bring your own OpenAI-compatible local server for MAI Code 1.1 Flash or Qwen3.8 Flash Next across Windows, macOS and Linux. Docs say those local models are more likely to be used on Fusion Standard when appropriate.

The gotcha is printed in plain language: local inference does not make your session offline. Fusion still makes routing decisions in the cloud, and local models may have different safety protections from Copilot-provided cloud models. Experimental features must be enabled (Copilot app “Local model auto routing”, CLI --experimental or /experimental on, VS Code chat.copilot.localModel.enabled). Adding a BYOK model does not automatically make it eligible for Fusion.

Creator videos that title the feature as routing AI to “your own PC” are useful as demand signal and weak as a privacy summary. The primary pages are clearer: on-device weights can run for some Standard steps, while the router that decides when to use them stays on GitHub’s side of the wire.

Where it ships, and what we could not verify

GA surfaces named on Oct 28 are Visual Studio Code, Copilot CLI, the GitHub Copilot app and JetBrains IDEs. Support for additional clients is “coming soon” without a dated list. Research-preview users who had auto, HydraFusion or HydraFusion Max selected are told those choices transition to the matching Fusion effort levels.

Official video context from the same launch window includes GitHub’s Universe 2026 opening keynote and the day-1 keynote recap. We treat those as official launch media, not as independent benchmarks.

  • A exact public launch date for the earlier Project HydraFusion research preview. The GA changelog and docs name HydraFusion as lineage; we did not find a dated company changelog node for the preview start to put on this timeline.
  • Enterprise price quotes for a specific team. AA’s $5.59-per-task figure is a benchmark average under AA’s harness, not a GitHub price list for your org.
  • Hands-on proof from this page that Fusion Max will beat Claude Code on your repository. We did not run either agent.
  • An offline-capable Copilot Fusion session. Primary pages say routing decisions stay in the cloud even when a local model runs a step.

Our take

Front’s angle is narrower than the scoreboard posts. Fusion is a runtime policy you subscribe to: you choose how hard you want the system to try, GitHub chooses which models and which review pattern fire, and you pay the underlying meters with a temporary 30% haircut through November 30, 2026. If your admin has locked the interesting models out of Fusion’s pool, the picker entry will look emptier than the launch blog.

Keep model-family timelines separate when you need version history. The OpenAI GPT model release timeline and the Claude model release timeline still date the weights Fusion may call. For open-weight and storefront notes that sit beside coding tooling, see Puter and The Everything Project and GPTibo open source: what is verified. More dated notes live on the Front blog index.

Keep reading on Front

Every post, newest first, is on the Front blog index.

FAQ

Is GitHub Copilot Fusion a single model?

No. GitHub’s docs call it a runtime model router and orchestrator. Selecting Fusion chooses the system; Fusion then picks models and an execution pattern per prompt.

What do Fusion Standard, High and Max change?

They change the quality, cost and capability tradeoffs Fusion prioritizes. Standard favors versatile everyday models; High favors stronger models for harder tasks; Max favors the most capable models. Docs also describe Fusion Fast, which users cannot select and which appears only in VS Code inline chat.

How is Copilot Fusion billed?

You are billed for each underlying model Fusion uses, at that model’s standard rate. There is no separate Fusion fee. Through November 30, 2026, paid Copilot plans receive a 30% discount on those model costs when using Fusion, according to GitHub’s docs.

Does local-model routing make Copilot Fusion offline?

No. GitHub’s local-models changelog and docs say Fusion still makes routing decisions in the cloud. Local models are an experimental preview and do not by themselves create an offline session.

What did Artificial Analysis report for Fusion Max?

AA’s Oct 28 article reports Copilot Fusion Max at 60 on the Coding Agent Index versus 62 for Claude Code with Opus 5.5 Max, about $5.59 versus about $11.42 per task. Those are AA-reported figures, not a GitHub price sheet.

Update log

  • — First published. Checked GitHub’s Copilot Fusion GA changelog, the Copilot Fusion docs page (including billing and local-model notes), the Oct 28 local-models changelog, @github’s Oct 28 X post, and Artificial Analysis’s Coding Agent Index write-up as a labeled secondary source, through Oct 30, 2026 (UTC+8).

Sources