Agents

GitHub Copilot Fusion picks models for you: better coding for less?

Copilot Fusion selects and combines AI models. Its benchmark score, cost per task, limitations and what GitHub credits mean for developers.

By Audryk · October 29, 2026

• Coding agents
• Models
• #GitHub Copilot
• #agents
• #models
• #pricing
• #benchmarks

Should developers use a powerful model for every task, or save it for the hardest problems? Copilot Fusion, launched on October 28, offers another approach: let Copilot choose and, when needed, combine models for you (GitHub).

The promise is attractive: approach top coding-agent performance without paying for the most powerful model on every task. An independent evaluation lends it some support, but it does not prove Fusion will always be cheaper on your project. Here's what the numbers measure and how billing works.

An orchestrator, not a new model

Copilot Fusion routes and orchestrates models. Choose an effort level and Fusion selects from the models available to your plan, deciding how to use them. It may use one model or add review steps when a task warrants them (GitHub documentation).

GitHub describes three execution patterns:

Source: GitHub documentation. These are possible patterns; selection depends on the task and chosen effort.

Standard targets everyday work and, according to the documentation, uses one model per task. High targets substantial changes; Max prioritizes the most demanding problems. GitHub says High and Max can add model calls and use more credits. Users can follow the steps and inspect the models used beneath the response (GitHub).

An independent test: two points lower at about half the cost

Artificial Analysis evaluated Fusion Max in Copilot CLI on its Coding Agent Index v2.0. The index equally weights FrontierCode v1.1, Terminal-Bench 4.0 and SWE-Atlas-QnA-AA. The comparison is between an agent and model configuration, rather than isolated models (Artificial Analysis).

Source: Artificial Analysis agent comparison, checked October 29, 2026. These are costs calculated for benchmark tasks, not the price of a task under your subscription.

Fusion lands two points below Claude Code on this index, at roughly 51% lower average cost. Results vary by test: Fusion scores 71% on Terminal-Bench 4.0 versus Claude Code's 63%, but 51% on FrontierCode versus 56%. Similar averages do not mean both agents solve the same problems (Artificial Analysis).

The result cannot predict your savings. Benchmark tasks, configurations, tools, prices and cache conditions may differ from yours. It compares complete systems with their respective environments, and does not establish whether their output fits your architecture or is easy to maintain.

Billing: credits, not a flat Fusion rate

There is no separate Fusion fee. GitHub charges for each model used at its respective rate, so a task involving multiple models can cost more than a single call. On individual plans, token usage is converted into GitHub AI Credits; one credit represents $0.01 of usage. Included allowances and exact rules depend on the plan (GitHub individual billing).

GitHub lists these monthly allowances for paid individual plans:

Source: GitHub documentation, checked October 29, 2026. GitHub says part of the allowance is variable. Free, student and business plans have different conditions; check your usage dashboard before comparing.

Paid Copilot plans receive a 30% discount on model costs when using Fusion through November 30, 2026. This is not a permanent subscription discount (GitHub).

Start with Standard for a focused change. Move to High or Max when the difficulty warrants more reasoning, then review credits consumed and time spent checking the work. The meaningful cost is the cost of an accepted, verified solution, not just a first response.

Availability and local models

Fusion is available in VS Code, Copilot CLI, the GitHub Copilot app and JetBrains IDEs, subject to plan and organization policies. Update your client, select Copilot Fusion in the model picker, and choose Standard, High or Max (GitHub).

GitHub is also testing local model selection at the Standard effort level. In Copilot CLI, certain manually configured models can run on macOS, Windows and Linux; users must manage the setup and provide sufficient hardware. This is an experimental preview. Even if inference runs locally, routing remains in the cloud: the session is not fully offline (GitHub).

Our take

Fusion addresses a real problem: using the strongest model for a small task can waste credits, while a lightweight model can fail on a complex change. Orchestration makes that trade-off more automatic, but it does not remove the need to review changes or the uncertainty around final cost.

The Artificial Analysis result is promising for Fusion Max; it does not establish a win against every competitor on every task. A useful trial is to compare representative work: a small change, a difficult bug and a multi-file refactor, using the same success criteria while tracking credits and review time.

Sources

• GitHub, Copilot Fusion launch, October 28, 2026
• GitHub, Fusion operation, effort levels and billing, checked October 29, 2026
• Artificial Analysis, Copilot Fusion evaluation, October 28, 2026
• Artificial Analysis, Claude Code and Copilot comparison, checked October 29, 2026
• GitHub, individual billing and credits, checked October 29, 2026
• GitHub, local models in Copilot Fusion, October 28, 2026

Spotted an error in this article? Report it · Our corrections policy