GitHub Copilot Fusion is a setting in the Copilot model picker that chooses models for you on every prompt and, when a task looks hard enough, makes several models work together on it: one drafts, others review, and the first revises. It became generally available on October 28, 2026 at GitHub Universe, in VS Code, JetBrains IDEs, Copilot CLI and the GitHub Copilot app. There is no extra fee: you pay each model it calls at that model's normal AI credit rate, and paid plans get 30% off those model costs through November 30, 2026. Fusion Standard is on every plan, including Free; Fusion High and Fusion Max need Pro+, Max, Business or Enterprise.
This guide explains what Copilot Fusion does under the hood, what each effort level costs you in practice, how it differs from the older Auto option, and when picking a single model by hand is still the better call. Facts come from GitHub's Fusion announcement and Fusion documentation, checked on November 3, 2026.
What's new (November 19, 2026): GitHub added two Microsoft coding models to Copilot on November 17: MAI Code 2 for Pro, Pro+, Max, Business and Enterprise, and MAI Code 2 Flash, which also reaches Free and Student. Both are billed by usage. GitHub's models and pricing page lists promotional rates until May 17, 2027: $0.50 input and $2.50 output per million tokens for MAI Code 2, and $0.05 and $0.25 for the Flash model. MAI-Code-1.1-Flash will be deprecated in every Copilot experience on December 18, 2026, which matters for the local model option below. GitHub has not said which model replaces it on device.
Earlier (November 7 and 11, 2026): Copilot CLI 1.0.107 added a saved Fusion tier (copilot config fusionTier), and 1.0.109 added Claude Fable 5.2 to the model picker.
What Copilot Fusion actually does
GitHub describes Copilot Fusion as a runtime model router and orchestrator. When you select it, you are not choosing a model. You are choosing a system that reads each prompt, scores it on reasoning, code generation, debugging and tool use, and then picks both the models and the way they cooperate. It merges two earlier pieces: Project HydraFusion, a research preview that combined models on one task, and Copilot auto model selection, which picked one model per prompt.
Every prompt runs one of three execution patterns:
| Pattern | What happens | Cost effect |
|---|---|---|
| Single | One model does the task | Same as calling that model yourself |
| Cascade | A cheaper model drafts, a judge checks it, and the work escalates to a stronger model only if the draft fails | Extra judge call, sometimes a second model |
| Critique | One model drafts, several models critique it in parallel, and the first model revises using the combined feedback | Several extra calls |
The pattern is chosen per prompt, so a session can run a simple question as Single and the next, harder request as Critique. GitHub says choosing the pattern adds little delay, and that review calls are only added when the task warrants them. While it works you can follow each step and the model in use, but only the final answer is shown. After a response, hovering over the effort level name lists the models that took part.
Two details matter for long sessions. Fusion has no context window of its own: each step runs inside the limits of whichever model is doing it, and the window the picker shows is a cautious figure based on the smallest model in the pool. And Fusion tries to keep your main conversation on the same model so you keep the benefit of cached tokens, sending helper models only the context they need.
Fusion Standard vs High vs Max
You pick an effort level, and Fusion decides the rest. The plan table below comes from GitHub's documentation.
| Effort level | What GitHub says it prioritizes | Plans |
|---|---|---|
| Fusion Standard | Balance of cost, quality and speed; usually one model per task | All plans, Free and Student included |
| Fusion High | Orchestration with moderate token savings, for hard bugs or multifile changes | Pro+, Max, Business, Enterprise |
| Fusion Max | Frontier capability while still watching tokens | Pro+, Max, Business, Enterprise |
| Fusion Fast | Fast models only | Not selectable; used in VS Code inline chat |
In plain terms, Standard behaves most like a smart model picker, while High and Max are where the Cascade and Critique patterns earn their keep. Copilot Pro at $10 a month gets Standard only, which surprises people who assume every paid plan gets the full feature. To reach High and Max as an individual you need Pro+ at $39 a month or Max at $100 a month, per GitHub's plans page.
Copilot Fusion cost and the 30% discount
Fusion is billed by usage, not as a separate product. Every model it calls is charged at that model's standard per token rate and converted into GitHub AI Credits, where one credit is $0.01. Individual plans include a monthly allowance: 1,500 credits on Pro, 7,000 on Pro+ and 20,000 on Max (base plus flex). Business and Enterprise seats add 1,900 and 3,900 credits per user to a shared pool.
That billing model has a direct consequence. Fusion Standard on an easy prompt can be cheaper than you picking a premium model by hand, because it may route the prompt to a smaller one. Fusion Max on a hard prompt can cost more than a single premium model call, because Critique adds reviewer calls on top of the draft. GitHub's docs say this outright: High and Max may use more credits per task than Standard.
The launch discount takes 30% off model costs when you use Fusion on a paid plan, through November 30, 2026. Code completions and next edit suggestions stay unlimited on paid plans and never touch credits, so the discount only matters for chat, agent mode, the CLI and similar features.
If you want to try High or Max without risk of overage charges, GitHub's billing docs suggest setting a bundled AI credits budget of $0 with Stop usage enabled. You then spend only your included allowance. Run your usual tasks for a week on Standard, then a week on High, and compare the usage report.
Copilot Fusion vs Auto
This is the most common question in search right now, and the answer depends on where you use Copilot. Fusion did not delete Auto. GitHub's auto model selection page says that if you see Copilot Fusion in the picker instead of Auto, Fusion is available on that surface. Auto remains on GitHub.com chat and cloud agent, GitHub Mobile, Visual Studio, Xcode and Eclipse.
| Item | Copilot Fusion | Auto model selection |
|---|---|---|
| Where | VS Code, JetBrains, Copilot CLI, Copilot app | GitHub.com, Mobile, cloud, Visual Studio, Xcode, Eclipse |
| Models per prompt | One or several (Single, Cascade, Critique) | One |
| Settings | Standard, High, Max | Efficiency, Balance, Intelligence |
| Discount on paid plans | 30% through Nov 30, 2026 | 10% in Copilot Chat and cloud agent |
| Local models | Standard can use certain local models (preview) | Excludes local models |
Auto picks a single model per prompt and hands it the job. Fusion can do the same thing on Standard, but on High and Max it can stack a reviewer or escalate a draft. If you were in the HydraFusion preview or used Auto in a client that now has Fusion, GitHub moved your choice to the matching Fusion effort level automatically.
Visual Studio users are the main group still on Auto today. If that is you and you want other ways to get model choice, see our guide to the GitHub Copilot alternatives for Visual Studio.
How to turn on Copilot Fusion
Update your client first, because Fusion only appears in recent builds.
- VS Code and JetBrains: open the Copilot Chat model picker, choose Copilot Fusion, then the effort level.
- GitHub Copilot app: click the model under the prompt box, switch on Copilot Fusion and pick an effort level. Switch it off to choose a specific model again.
- Copilot CLI: type
/modelin a session and select Copilot Fusion, or set a default with/config model. GitHub's CLI docs give this command for scripts:
copilot --model fusion --fusion-tier standard -p "Fix the failing test in utils"
On Copilot CLI 1.0.105 that exact flag was refused. Version 1.0.107 (November 6, 2026) added a Fusion tier setting, and in our test it accepted both the flag and a saved default: copilot config fusionTier high writes the tier to your settings file, and anything other than standard, high or max is rejected. The flag still does not appear in copilot --help, and a managed organization setting can override your saved tier. Update with copilot update, sign in, and use /model inside a session if your account does not offer the tier you picked.
copilot config fusionTier high
If Fusion does not appear at all, your organization may have disabled every model it would use. GitHub says Fusion only draws on models enabled by your plan and your administrator's policies, and hides itself when none are available.
Local models in Fusion Standard
In public preview, Fusion Standard can route some work to a model running on your own machine, which uses no AI credits. On Windows with suitable hardware you can install MAI Code 1.1 Flash from the model picker in the CLI, the app or VS Code, until its December 18 deprecation. In Copilot CLI on Windows, macOS or Linux, Fusion can also use Qwen3.8 Flash Next served locally, for example through Ollama, if Copilot identifies it as qwen3.8-flash-next-local. You must turn on experimental features first: --experimental or /experimental on in the CLI, or Local model auto routing in the app's settings.
What the benchmark claim does and does not tell you
GitHub's headline number is that Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code running Opus 5.5 Max, at about half the average API cost per task, about half the tokens and roughly 40% less time. That is a useful signal that routing and review can close most of the gap to a top single model for less money.
It is still one index, run on someone else's tasks, and the cost comparison uses API prices rather than what you pay inside a Copilot plan or a Claude subscription. Your repository, your prompts and how often your tasks trigger Critique will decide your real numbers. Treat the claim as a reason to run your own two week comparison, not as a verdict. Our GitHub Copilot vs Claude Code guide covers the wider differences between the two products, from form factor to billing.
Limits to know before you rely on it
GitHub lists three limitations, and the second one deserves attention in agent mode.
- Fusion works separately from subagents and does not start them. If you delegate work to a custom agent pinned to a model, that agent keeps its own model.
- If Fusion throws away a draft, edits that draft already made to your files are not undone automatically. Review the diff before you commit.
- The execution patterns and the model pool will change over time, so results on the same prompt may shift from month to month.
For enterprises, administrators can see which models Fusion picks from, disable a model for Fusion without disabling it for everyone, and make any Fusion level the default through Enterprise managed settings. Data residency and FedRAMP policies also filter the pool.
Which setting should you use
For most people on Pro, Fusion Standard is a sensible default: it costs about what a good single model would and saves you from leaving an expensive model selected for trivial prompts, which is the problem an AT&T engineer described in GitHub's announcement. On Pro+ or higher, keep Standard as the default and switch to High for a multifile change or a bug you have already failed to fix once. Save Max for design questions and hard debugging where a second opinion is worth paying for.
Pick a specific model by hand when you need repeatable output for a benchmark or a CI job, when a task depends on one model's large context window, or when your team has standardized on one model for review. If you want this kind of routing outside Copilot, for your own agents, an LLM gateway or OpenRouter gives you rules you control. And if you are still choosing a coding assistant at all, start with our best AI coding assistant roundup. If you also connect outside tools to GitHub Copilot, our guide on how to use an MCP server in GitHub Copilot shows where each client reads its config.
FAQ
Is Copilot Fusion free?
Fusion Standard is available on Copilot Free and Copilot Student, and it draws on the plan's limited credit allowance. Fusion itself has no fee on any plan: you pay only for the model calls it makes, at standard rates.
What is the difference between Copilot Fusion and Auto?
Auto picks one model per prompt. Fusion can pick one model or combine several through Cascade and Critique, offers Standard, High and Max levels, and replaced Auto in VS Code, JetBrains, Copilot CLI and the Copilot app. Auto remains on GitHub.com, Mobile, Visual Studio, Xcode and Eclipse.
How long does the Copilot Fusion discount last?
GitHub's documentation says users on paid Copilot plans get 30% off model costs when using Fusion through November 30, 2026. Auto keeps its separate 10% discount on paid plans.
Can I see which models Copilot Fusion used?
Yes. After a response, hover over the effort level name below it to see the list of models. You can also follow each step and its model while Fusion is working.
What happened to HydraFusion?
Project HydraFusion was the research preview that became Copilot Fusion. If you had enabled Auto, HydraFusion or HydraFusion Max, GitHub moved those selections to the matching Fusion effort levels.

