GitHub Copilot Fusion: Standard vs High vs Max, Billing and BYOK
Use Fusion Standard for routine coding, High for bounded complex changes, and Max when deeper reasoning warrants extra usage. Learn which plans include each tier, how the 30% model-cost discount works through November 30, 2026, why BYOK models do not automatically join Fusion, and how to control enterprise defaults.
Contents
Use GitHub Copilot Fusion Standard for routine coding when you want Copilot to choose the model. Move to High for a difficult, bounded change, and test Max when extra reasoning could reduce costly rework. Select a model directly when you need a fixed model or your own API endpoint.
As of 2026-10-31, Fusion is generally available in VS Code, Copilot CLI, the GitHub Copilot app, and JetBrains IDEs. GitHub’s launch announcement describes a managed system for choosing and coordinating models. Your plan and administrator policies still determine what it can use.
What does GitHub Copilot Fusion actually do?
Fusion chooses how to solve each request, including which models participate. You select an effort level instead of choosing the underlying model yourself.
The Fusion documentation describes three execution patterns:
| Pattern | What happens | Practical consequence |
|---|---|---|
| Single | One model handles the task. | No separate model review stage is added. |
| Cascade | An efficient model drafts a solution. A judge accepts it or escalates to a stronger model. | A difficult request can incur drafting, judging, and escalation work. |
| Critique | Other models review a draft in parallel. The original model revises it using their combined feedback. | Review and revision add model usage and time. |
The pattern can change between prompts in one conversation. High or Max does not force every request through every stage.
Artificial Analysis’s launch report named Claude Opus 5.5, Claude Sonnet 5.5, GPT-6 Astra, and GPT-6 Sol among the models. Treat that as a launch snapshot. GitHub can change the pool, and your policies can reduce it.
Standard vs High vs Max: which can you use?
Start with Standard unless a specific task justifies more reasoning or review. The current plan table is narrower than the phrase “available on all plans” suggests.
| Effort | Recommended starting point | Multi-model orchestration | Eligible plans as of 2026-10-31 |
|---|---|---|---|
| Fusion Standard | Routine coding, explanations, and straightforward changes | Selects one model for the task | Free, Student, Pro, Pro+, Max, Business, Enterprise |
| Fusion High | Complex bugs or changes across several files with clear acceptance criteria | Can add model calls for review or escalation | Pro+, Max, Business, Enterprise |
| Fusion Max | Difficult investigations where a missed issue would cause expensive rework | Can add model calls; prioritizes deep reasoning | Pro+, Max, Business, Enterprise |
Copilot Pro does not include Fusion High or Max. The VS Code 1.144 release notes independently confirm the eligible paid tiers.
Standard’s “one model” does not mean one API call for an entire coding task. An agent can call the same model repeatedly while reading files, using tools, and checking results.
Fusion Fast is a separate option used internally for VS Code inline chat. Users cannot select it.
Choose High before Max when the change is substantial but well defined. Move to Max if High misses important dependencies or cannot explain a failure. These are starting recommendations, not measured performance guarantees for either tier.
How to try Fusion and inspect the result
Use an updated supported client and begin with a small repository task. In an IDE or the Copilot app, open the model picker, select Copilot Fusion, and choose Standard.
For a CLI trial, use GitHub’s installation instructions. The npm route requires Node.js 22 or later. Run this in Bash from your repository:
npm install -g @github/copilot
copilot --model fusion --fusion-tier standardIf prompted, enter /login and complete GitHub authentication. Organization-provided access also requires the Copilot CLI policy to be enabled.
Enter this prompt in the interactive session:
Trace how this repository loads and validates configuration.
Identify the relevant source files and existing tests.
Propose one missing edge-case test and explain its expected result.
Do not edit files.A useful result identifies real files and connects the proposed test to existing behavior. Check those references against the code. The answer appears in the same session.
The CLI model-selection reference documents --model fusion and --fusion-tier standard. Use high or max for another eligible tier. In an interactive session, /model opens the picker.
Check the selected model in the CLI status line. In VS Code, hover over the effort label below the response to inspect the models used. A simple task returning through one model is normal.
If Fusion is missing, update the client and check plan eligibility and model policies. Fusion disappears when none of its eligible models are available. Also check whether you remain on a legacy annual request-based plan, which does not receive new models and features.
How does Fusion consume AI credits?
Fusion charges for the underlying model usage, with no separate orchestration fee. High and Max can consume more credits because extra drafting, judging, or revision steps generate additional billable tokens.
GitHub’s model pricing reference prices input, output, cached input, and applicable cache writes separately. One AI credit represents $0.01 USD of usage. One prompt therefore has no fixed credit price.
As of 2026-10-31, GitHub’s billing documentation offers 30% off Fusion model costs through November 30, 2026. Under its paid-plan wording, paid Business and Enterprise subscriptions qualify. The discount concerns model usage; seat subscription charges remain separate.
For an illustrative calculation, assume all model calls for one task total $1.00 at standard rates, including applicable cache charges:
- Eligible promotional usage: $1.00 × 0.70 = $0.70, or 70 AI credits.
- The same usage at undiscounted rates: $1.00, or 100 AI credits.
These are assumed totals, not a prediction for the example prompt. For an actual estimate, add the token charges for every participating model.
Credits drawn from an included allowance reduce that allowance. They become additional paid usage only after the allowance is exhausted and paid usage is enabled.
On Business and Enterprise, usage consumes the shared included-credit pool first. Additional charges follow when that pool is exhausted and paid usage is allowed. Administrators who want no overage must disable AI credits paid usage; it is enabled by default.
After a trial, inspect AI usage under GitHub billing settings. Compare consumption by model with the quality of the result. Fusion tries to preserve the main model’s cache, but that does not guarantee lower costs for every task.
Will Fusion use your BYOK endpoint or local model?
Adding a model through bring your own key, or BYOK, does not automatically make it eligible for Fusion. GitHub states this explicitly in the local-model routing announcement.
With user-configured BYOK, a supported client sends model requests to your configured provider. Select that custom model when you require a particular endpoint or provider account. Its appearance in the picker is not proof that Fusion can route to it.
For a separate compatible API workflow, BetterToken documents the OpenAI-compatible base URL https://www.bettertoken.ai/v1. Configure the protocol required by your client. Adding a gateway endpoint does not itself enroll its models in Fusion.
Local routing is a specific preview exception:
- Fusion Standard can use supported local MAI Code 1.1 Flash installations. Managed downloads use Windows ML on supported Windows devices.
- Copilot CLI also supports eligible manually configured local models on Windows, macOS, and Linux. GitHub documents Qwen3.8 Flash Next recognition as
qwen3.8-flash-next-local. - In the CLI, enable experimental features with
--experimentalor/experimental on, then choose Standard. In the Copilot app, automatic local routing currently supports managed Windows ML models.
Follow GitHub’s local-model setup instructions for provider configuration and hardware requirements. Fusion still makes routing decisions in the cloud, so local inference does not make Fusion an offline workflow.
If you want to select a local model explicitly in another coding tool, see using a local model in Claude Code and switching back.
Fusion Max vs Claude Code: what does the benchmark show?
Fusion Max is worth testing when metered API cost matters, but the published result does not establish a cheaper subscription. Artificial Analysis tested Max in Copilot CLI, not Standard or High.
Its October 28 report and current comparison, checked on 2026-10-31, show:
| Metric | Copilot CLI with Fusion Max | Claude Code with Opus 5.5 Max, Default Fallback |
|---|---|---|
| Coding Agent Index v2.0 | 60 | 62 |
| Average API cost per task, USD | $5.59 | $11.42 |
| Terminal-Bench 4.0 | 71% | 63% |
| FrontierCode v1.1 | 51% | 56% |
| SWE-Atlas-QnA-AA | 57% | 66% |
The reported cost difference is $11.42 − $5.59 = $5.83 per benchmark task. That is (1 − 5.59 ÷ 11.42) × 100 = approximately 51.1% lower reported benchmark API cost.
AA’s methodology accounts for applicable cached-input and cache-write prices. Its cost metric measures token usage, not consumer subscriptions or full operating costs.
Do not apply another 30% reduction to $5.59. The cited AA pages do not establish how GitHub’s temporary promotion is reflected in that figure.
My recommendation from these results is to test Fusion Max first for difficult terminal work when API usage cost matters. Test the Claude Code variant first when repository-change or codebase-question accuracy outweighs the reported cost gap. Reverse that choice if your own task results, repair time, or billing differ.
For a comparable alternative, configure Claude Code to use Opus 5.5 and record the effort and fallback settings. The published “Default Fallback” qualifier is part of the tested configuration.
How can administrators control Fusion?
Control the allowed models separately from the default selection. A default does not replace model policies or spending limits.
Enterprise owners can open AI controls → Copilot. Under Models, click Configure Copilot Fusion models. Set each model’s status there, following the enterprise model controls. This can exclude a model from Fusion while leaving it available for manual selection.
Other model policies still apply. Fusion cannot bypass plan restrictions, organization restrictions, or policies for data residency, FedRAMP compliance, and evaluation models.
For a GitHub-hosted enterprise configuration:
- Create or use a designated organization’s
.github-privaterepository. - Select that organization under AI controls → Agents → Configuration source.
- Add the following JSON at
copilot/managed-settings.json. - Commit it to the repository’s default branch.
{
"model": "fusion",
"fusionTier": "standard"
}This example makes Fusion the default for new conversations and locks its effort to Standard. Users can still select another model per conversation. A plain fusionTier string prevents user and repository overrides; the managed-settings reference documents overridable for enterprise-team exceptions.
The fusionTier key requires Copilot CLI 1.0.104-3, VS Code 1.144.0, or Copilot app 1.1.33-4, or later. As of 2026-10-31, JetBrains supports Fusion but does not support these managed model and fusionTier keys.
Follow the managed-settings rollout guide to check validation issues under AI controls → Agents. Restart the client or sign in again to refresh settings, then verify a new conversation starts in Fusion Standard.
If the default does not appear, check client version, the selected configuration source, and which enterprise provides the user’s Copilot license.
Which limitations should change your workflow?
Review workspace changes even when Fusion rejects an intermediate draft. GitHub documents that file edits already made by a discarded draft are not automatically undone.
After an editing task, inspect git status, git diff, and git diff --cached. Review new files and run the relevant checks before committing. A polished final response does not prove every earlier edit remains appropriate.
Fusion also has no context window of its own. Each step uses its model’s context limit, and the displayed Fusion limit reflects a conservative value. Switching to Max does not combine several models’ context windows.
Finally, Fusion does not start subagents. If your workflow depends on separate agents with explicit roles, configure that delegation separately. Fusion’s internal reviewers do not replace those agent instructions.
