On October 28, 2026, GitHub announced the general availability of GitHub Copilot Fusion, a feature that uses multiple AI models to assist with coding work. It automatically selects an appropriate model for each task and, when needed, adds judging or review by other models. Fusion combines the existing automatic model selection feature with technology from Project HydraFusion, GitHub's research into multi-model coordination. It is available in VS Code, the GitHub Copilot app, Copilot CLI, and JetBrains IDEs.
In a benchmark by independent evaluator Artificial Analysis, Copilot Fusion run at the Max setting in Copilot CLI recorded an overall score of 60 and an average task cost of $5.59. By comparison, Claude Code with Opus 5.5 Max scored 62 overall at an average of $11.42. Fusion came within 2 points on the overall score while cutting the average task cost by roughly half. However, this result does not mean the same quality or cost savings will apply to every kind of development work. What distinguishes Copilot Fusion is that it does more than switch models: it automatically decides whether to add judging or review depending on the task.
Overall score of 60 in independent evaluation, with average task cost down about 51%
Artificial Analysis's evaluation, published on October 28, compares the performance and cost of coding agents, including their execution environments, rather than of AI models alone. The "Coding Agent Index v2.0" calculates an overall score by combining three benchmarks with equal weighting. Copilot Fusion Max scored 60 overall, while Claude Code with Opus 5.5 Max and Devin Fusion each scored 62. Note that this overall score is a composite of multiple evaluation results, so a score of 60 cannot be read as a 60% task success rate.
On cost, Copilot Fusion's average task cost was $5.59, compared with $11.42 for Claude Code with Opus 5.5 Max and $8.00 for Devin Fusion. Using Claude Code as the baseline, Copilot Fusion's average task cost is about 51.1% lower. However, this is an average based on API costs in the benchmark. It does not represent what actual Copilot users pay, nor the cost per successfully completed task. It also excludes monthly product fees and the cost of human code review, so it cannot be directly translated into a reduction in overall development costs.
Looking at the individual benchmarks, Copilot Fusion scored 71 on Terminal-Bench 4.0, 51 on FrontierCode, and 57 on SWE-Atlas-QnA-AA. It took first place among the compared products on Terminal-Bench 4.0, but on FrontierCode, Claude Code with Opus 5.5 Max (56) and Devin Fusion (54) scored higher than Copilot Fusion. Even when overall scores are close, the strengths of each product differ by benchmark. In actual adoption, it is necessary to verify performance on tasks close to your own development work.
Token consumption also differed significantly. According to Artificial Analysis, Copilot Fusion used an average of 7.2 million tokens per task, about half the amount used by Claude Code with Opus 5.5. However, this comparison alone does not reveal how much individual mechanisms, such as automatic model selection, cache utilization, and multi-model coordination, contributed to the cost reduction. These figures should be understood as evaluation results for the agent's entire execution approach, not for the performance or unit price of any single model.
Automatically choosing between a single model and judging or review by other models
Project HydraFusion, the foundation of Copilot Fusion, was announced as a research preview on September 4, 2026. Instead of leaving all processing to a single model, it combines drafting an initial answer, quality judging, review, revision, and handing off to a higher-performance model, depending on the task. With general availability, this multi-model coordination has been integrated with the existing automatic model selection and can be used from the regular Copilot environment.
According to GitHub's official documentation, Copilot Fusion switches among three execution patterns: Single, Cascade, and Critique. Single completes processing with one model. In Cascade, the draft produced by the first model is judged and, if necessary, handed off to a higher-performance model. In Critique, multiple other models independently review the draft in parallel, and the first model revises it based on the aggregated results. The differences between the patterns are summarized below.
| Execution pattern | How it works | Impact on calls and wait time |
|---|---|---|
| Single | One model handles the task | No additional review step |
| Cascade | Judges the initial draft and hands off to a higher-performance model if needed | Additional calls, such as for judging, occur |
| Critique | Multiple models review in parallel, and the original model revises its answer | Review and revision steps are added |
Single keeps processing simple by skipping extra review, while Cascade and Critique make additional model calls only when quality improvement is expected. This is designed to avoid unnecessary computation and waiting for simple tasks while adding the necessary processing for complex ones. However, additional judging and review cost both money and time, so using multiple models is not more efficient for every task.
The "processing level" that users select is a separate concept from these execution patterns. Copilot Fusion has three main processing levels: Standard, High, and Max. Standard emphasizes a balance of quality, cost, and wait time for everyday work, choosing one model per task. High is intended for complex but well-scoped development tasks, and Max prioritizes the most advanced reasoning and analysis. High and Max are not tied to a specific execution pattern; even within the same session, different approaches may be chosen depending on the request.
Copilot Fusion itself is available on all Copilot plans, but the selectable processing levels vary by plan. Standard is available on all plans, while High and Max are limited to supported plans. The available AI models are also restricted by subscription plan and organizational policy, so not every user can run processing with the same model lineup.
In Artificial Analysis's October 28 evaluation, Copilot Fusion's candidate models included Anthropic's Claude Opus 5.5 and Claude Sonnet 5.5, and OpenAI's GPT-6 Astra and GPT-6 Sol. Copilot Fusion chooses from these models according to the nature of the task and coordinates multiple models as needed. However, the candidate lineup is not fixed and will change as new models are added and according to GitHub's evaluation results.
When using multiple models, not only token consumption but also sharing of conversation history and reuse of caches affect cost. According to GitHub, Copilot Fusion keeps using the same model in the main conversation wherever possible so that cached tokens can be reused. Meanwhile, auxiliary models responsible for tasks such as review are given only the necessary context to avoid wasteful input. When switching models, it also takes into account cache boundaries and changes in conversation complexity, avoiding additional costs that do not improve quality, GitHub says.
These measures are an important mechanism for keeping costs down while using multiple models. However, it is not clear how much cache efficiency, model selection, and the review process each contributed to the roughly 51.1% difference in average task cost confirmed in the independent evaluation. The benchmark result of lower cost should be distinguished from the cost savings actually achievable in real work.
The 51.1% lower average task cost is separate from the limited-time 30% discount
There is no additional charge for using Copilot Fusion itself. However, the standard rates of each AI model used internally apply. With High and Max in particular, multiple models may be used for a single task, so they may consume more AI credits than Standard. Additional model calls for review and quality judging add cost as well as processing time.
GitHub is running a campaign for paid Copilot plan users through November 30, 2026, offering a 30% discount on the cost of models used in Copilot Fusion. However, this 30% discount and the roughly 51.1% reduction in average task cost shown in the Artificial Analysis benchmark are figures obtained under entirely different conditions. They cannot simply be added or multiplied to calculate an actual reduction in usage fees. To judge cost-effectiveness, you need to check the AI credits actually consumed at each processing level and the amount of processing your development work requires.
Model selection is automatic, but checking file changes is still essential
The models available to Copilot Fusion are determined by your subscription plan and your organization's administrative policies. According to GitHub's announcement, administrators of enterprises and organizations can restrict the models Copilot Fusion can use, and can also disable specific models for Fusion only. In other words, automating model selection does not mean models the organization hasn't permitted will be used without approval. At adoption, you should check which models are selection candidates.
There are also constraints on context length. Copilot Fusion does not have its own context window; each process runs within the context limit of the model actually used. In addition, Fusion's multi-model coordination is a separate mechanism from subagents, and Fusion itself does not launch subagents. It is important not to confuse multi-model review with work done by independent subagents.
Another point to watch is the handling of generated code and file changes. GitHub's official documentation states that even if Copilot Fusion discards an intermediate draft, file edits that draft has already made to the workspace are not automatically undone. This means changes from processing that wasn't adopted in the final answer may remain in the working directory. Even when multiple models perform judging and review, the final state of the files is not automatically guaranteed, so reviewing the diff before committing is essential.
By automating not only AI model selection but also review and handoff to higher-performance models, Copilot Fusion aims to balance quality and cost in a coding agent. In the independent evaluation, it kept an overall score close to a competing product while cutting average task cost by about half, but that advantage does not necessarily hold for all development work. In actual adoption, the key considerations are whether it delivers the quality you need, the credit consumption and wait time at each processing level, and whether you can properly review the generated file changes.