GitHub has made Copilot Fusion generally available, giving every Copilot plan a mode that chooses and combines AI models for each coding task instead of leaving developers to pick one. The Fusion changelog post, published on 28 October during GitHub Universe, says it works in VS Code, the Copilot app, Copilot CLI and JetBrains IDEs, with more clients to follow.
GitHub also put two agent features into public preview in its Copilot app: Projects, a single home for work that spans several repositories, and computer use, which lets Copilot operate desktop apps on macOS and Windows. Taken together, the releases push Copilot further from autocomplete and towards a manager of several agents and models at once.
In brief
- Copilot Fusion routes each task to one or more models and is now generally available on all Copilot plans.
- GitHub says Fusion Max came within two points of Claude Code with Opus 5.5 Max on one benchmark, at about half the cost per task.
- Paid plans get 30% off model costs in Fusion until 30 November, according to GitHub.
- Copilot Projects and computer use are in public preview; computer use asks before it controls an app.
How Copilot Fusion picks its models
Fusion brings together two earlier pieces of work: GitHub’s Project HydraFusion, which orchestrated several models on one job, and Copilot’s automatic model selection. Users choose an effort level, Fusion Standard, High or Max, and Fusion decides which models to use.
It can run a task in three ways. In Single mode, one model does the work. In Cascade, an efficient model drafts the answer and a quality check either accepts it or escalates to a stronger model. In Critique, one model drafts, a second reviews, and the first revises.
GitHub’s headline claim is about cost. “In the Artificial Analysis Coding Agent Index, Copilot Fusion Max scored 60, within two points of Claude Code with Opus 5.5 Max,” the changelog says. “It delivered that result at about half the average API cost per task, using about half as many tokens and taking about 40% less time per task.” Artificial Analysis, which runs the index, said in its own post on X that Fusion scores two points below Opus 5.5 in Claude Code at half the cost per task.
AT&T is GitHub’s named early customer. Hari Pokhrel, a principal software engineer there, said Fusion Standard “has improved token efficiency across AT&T by automatically selecting the right model for the task”.
What it costs and what admins control
Fusion is billed for each model it calls, at that model’s standard rate. The changelog says paid Copilot users get “a 30% discount on model costs when using Copilot Fusion” for a limited time, and GitHub said on X that the offer runs until 30 November.
For businesses, Fusion only uses models already enabled in enterprise or organisation policies. Administrators can block particular models from Fusion alone and can make Fusion the default through enterprise managed settings. Users who chose auto, HydraFusion or HydraFusion Max during the research preview are moved across automatically. That last point matters for cost owners: teams already on the preview will start being billed under Fusion’s per-model rules without changing anything.
Projects and computer use enter preview
The Copilot app changelog describes Projects as a way to keep multi-repository work in one place, with context such as repositories, design mocks, Slack threads and meeting notes. Copilot runs “as many agents as needed” and returns results such as pull requests. Projects can also run automations, and the app now connects to Slack and Figma.
Computer use lets Copilot read what is on screen, click controls, type and move between apps, including software with no API or command-line access. GitHub says “Copilot asks for your approval before controlling an app, and your organization can disable the feature through managed settings.” It is switched on under Settings, then Computer Use, and on a Mac needs Accessibility and Screen Recording permissions. Rival Cursor gave its agents the same kind of desktop control last week.
A third post, on GitHub intelligence, makes Copilot Memory generally available on paid plans and puts “dreaming”, which reviews past sessions to update memories and propose changes to instruction files, into opt-in public preview. Dream runs consume AI credits.
Fusion’s savings rest on one benchmark and GitHub’s own sums
The cost claim is the reason most teams will try Fusion, and it deserves a careful reading. The figures come from a single index, and the comparison is with one rival setup, Claude Code running Opus 5.5 at its highest effort. Artificial Analysis’s post backs the headline numbers, but neither GitHub nor TSN has shown how Fusion performs on a company’s own code, where task mix and repository size can change both quality and cost. TSN has not tested Fusion.
Billing also becomes harder to predict. A Cascade run may stop at a cheap model or escalate to an expensive one, and a Critique run calls at least two models. That flexibility is where the savings come from, but it means the bill depends on decisions the user no longer makes. The 30% discount will flatter early costs until 30 November, so teams comparing Fusion with a fixed model should look at their spending after it ends.
Computer use raises a different question: what an agent can do on a desktop full of email, chat and company systems. Approval prompts and an organisation-wide switch are sensible defaults, but administrators should decide whether to allow the preview before staff turn it on.
A busier field for coding agents
Fusion extends the multi-model approach GitHub has used as it added models such as Gemini 4 Argon to Copilot this month. Competition is moving fast: Cursor has released an Agent SDK for building agents as code projects, and former GitHub chief executive Thomas Dohmke’s start-up Entire has launched a Git forge built for AI agents.
The next date to watch is 30 November, when Fusion’s launch discount ends and teams see what model routing really costs.
On the same day, Microsoft began offering developers the Copilot harness behind GitHub Copilot through Foundry, in public preview.
Update, 28 Oct 2026: Agentic Workflows reaches general availability and Copilot gains image generation
GitHub made two further announcements on 28 October. The first is that GitHub Agentic Workflows is now generally available on github.com for GitHub Free, Team and Enterprise. The open-source project from GitHub Next lets teams describe a task such as issue triage, CI failure analysis, documentation updates or test writing in a plain-language Markdown file, which is compiled into standard GitHub Actions YAML. Because the result is an Actions workflow, GitHub says it reuses existing runner groups and policy rules, and each workflow can run on Copilot or on another coding agent. Using the default Copilot CLI engine requires a Copilot licence.
GitHub says agents run with read-only permissions by default, inside a sandboxed container behind what it calls the Agent Workflow Firewall, and that a separate threat-detection job scans proposed changes before they are applied. New since the public preview are support for self-hosted Actions Runner Controller runners with Docker-in-Docker, a gh aw forecast command for estimating AI credit spend that is no longer experimental, a choice of model provider per workflow, a model router, and the option to run agents in a Cloud Hypervisor or NVX micro-VM.
GitHub’s announcement quotes Ben Alton of Accenture, who says: “One of our leads saved an estimated 29 hours a week on his development work alone.” He adds that “AI now authors 91% of the team’s commits” and that the team “cut production errors from about 11,000 a day to under 1,000”. Those are Accenture’s figures, quoted by GitHub, and have not been independently checked.
The second announcement puts image generation and editing into Copilot in public preview. When a user asks for an image, the selected reasoning model calls a tool powered by GPT Image 2.5 Flare, previews the result inline and saves it to the project. It works in VS Code, the GitHub Copilot app and Copilot CLI, for Copilot Pro, Pro+, Business and Enterprise. Images are billed separately from the chat model, on image tokens that count towards a plan’s AI credits. Organisation and enterprise owners can switch the feature off through Copilot model policies. A visible “Made with AI” watermark is available, but it is off by default.
Agents move from the editor into the pipeline
Taken with Fusion, these launches push Copilot beyond a chat box in the editor. Agentic Workflows puts agents inside the CI system that already runs a team’s tests and releases, which is where cost and security controls have to hold. That may explain why GitHub leads with read-only defaults, a firewall, a threat-detection pass and a credit forecast rather than with speed. Like Replit’s cloud agents that open tested pull requests, the bet is that teams will hand over routine work if every change still arrives for review. Each automated run also spends AI credits, so the forecast tool matters as much as the model router once Fusion’s discount ends on 30 November. Image generation follows the same metering, and its default settings deserve attention: the watermark is opt-in, so images made in Copilot will carry no visible label unless someone turns it on. Microsoft, which owns GitHub, is also building image models of its own, such as MAI Image 2.7, now in preview for editing.
Sources
- GitHub Changelog, “Route and orchestrate models with GitHub Copilot Fusion”, 28 October 2026
- GitHub Changelog, GitHub Copilot app Universe releases, 28 October 2026
- GitHub Changelog, “GitHub intelligence puts your code, context, and history to work”, 28 October 2026
- GitHub on X, 28 October 2026
- Artificial Analysis on X, 28 October 2026
- GitHub Docs, Copilot Fusion
- GitHub Changelog, “GitHub Agentic Workflows generally available”, 28 October 2026
- GitHub Changelog, “Generate and edit images directly in Copilot”, 28 October 2026
Get TSN’s biggest tech stories, explained.
News and insights across AI, tech, crypto and cybersecurity. Every source linked. Unsubscribe anytime.
Get the newsletterRelated stories
- GitHub Makes Slack and Teams Copilot Cloud Agent Generally Available
- Gemini 4 Argon Is Now Generally Available in GitHub Copilot: Who Gets It, What It Costs and What Admins Need to Do
- Cursor Agents Can Now Drive Desktop Apps; Macs Need 2 Permissions
- Cursor’s Agent SDK Turns AI Agents Into Code Projects You Can Test
- Microsoft Opens Its Copilot Agent Harness to Developers via Foundry
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X





