News GitHub
Copilot Fusion: close to Claude Code at half the task cost
Cover adapted by HeyMark using official artwork from GitHub Universe 2026. Charts: Artificial Analysis data, October 28.
GitHub has put Anthropic and OpenAI models on the same team. Its new Copilot Fusion decides which model to use at each step and can combine several when a task calls for it. In Artificial Analysis’ Coding Agent Index v2.0, Fusion Max finished two points behind Claude Code running Opus 5.5 Max, at about half the average API cost per task. The interesting part is the trade-off: Fusion leads the terminal test, while Opus keeps an advantage on the other two components. Artificial Analysis evaluation.
From HydraFusion preview to general availability
Announced at GitHub Universe on October 28, Copilot Fusion is now generally available. It brings together two earlier efforts: Project HydraFusion’s multi-model orchestration and Copilot’s automatic model selection. GitHub describes Fusion as a system that evaluates a task’s needs across reasoning, code generation, debugging, and tool use, then chooses how to run it. Developers choose an effort level; Fusion selects the models and execution steps. “Fusion” is not the name of one model. GitHub Universe announcement · Fusion documentation.
HydraFusion had already been available as a research preview. Today’s announcement is the general availability of that orchestration work under a new name, with automatic model selection integrated, rather than GitHub’s first orchestration experiment. GitHub says the pool can include Claude and OpenAI models. It can change over time and depends on the models allowed by a user’s plan and organization policies.
Copilot Fusion vs Claude Code: the benchmarks
Artificial Analysis evaluated Copilot Fusion Max in GitHub Copilot CLI using its Coding Agent Index v2.0. The comparison configuration was Claude Code with Opus 5.5 at Max effort; the cost chart labels it Default Fallback. The index gives equal weight to three benchmarks, so its headline score summarizes different kinds of work; it is neither a universal quality rating nor a measure of every developer’s day-to-day use. Artificial Analysis reports a score of 60 for Fusion and 62 for Claude Code, with average API cost per task of $5.59 and $11.42, respectively. Fusion used 7.2 million tokens per task, roughly half the token volume of Claude Code in this evaluation. (Artificial Analysis, October 28)
If you usually default to one model, look at the tests closest to your work. The average brings together quite different results.
What you gain, and where you trade off
Artificial Analysis · Coding Agent Index v2.0 · Oct 28, 2026
Composite index · higher is better
Coding Agent Index · higher is better
The overview weights the three benchmarks equally. Switch tests to see where each system leads.
GitHub Copilot · Copilot Fusion (max)
- Index
- 60
- USD / task
- 5.59
- FrontierCode v1.1
- 51%
- Terminal-Bench 4.0
- 71%
- SWE-Atlas-QnA-AA
- 57%
| Agent and configuration | FrontierCode v1.1 | Terminal-Bench 4.0 | SWE-Atlas-QnA-AA | Index | USD / task |
|---|---|---|---|---|---|
| Claude CodeClaude Opus 5.5 (max, Default Fallback) | 56% | 63% | 66% | 62 | 11.42 |
| Devin Fusion CLIClaude Opus 5.5 (max) + GPT-6 Luna (high) | 54% | 63% | 68% | 62 | 8.00 |
| GitHub CopilotCopilot Fusion (max) | 51% | 71% | 57% | 60 | 5.59 |
| Claude CodeClaude Sonnet 5.5 (max) | 48% | 66% | 64% | 59 | — |
| Muse CodeMuse Wonder Full (max) | 49% | 59% | 69% | 59 | — |
| Claude CodeClaude Fable 5.1 (max) | 49% | 58% | 66% | 58 | — |
| CodexGPT-6 Astra (max) | 49% | 56% | 59% | 55 | — |
| CodexGPT-6.1 Sol (max) | 49% | 53% | 60% | 54 | — |
| Antigravity CLIGemini 4 Argon (high) | 49% | 56% | 55% | 53 | — |
| Muse CodeMuse Wonder (max) | 46% | 48% | 62% | 52 | — |
| OpenCodeGLM-5.3 (max) | 48% | 40% | 59% | 49 | — |
| CodexGPT-6.1 Terra (xhigh) | 45% | 46% | 54% | 48 | — |
| Grok BuildGrok 4.7 (xhigh) | 44% | 33% | 63% | 47 | — |
| DeepSeek HarnessDeepSeek V4.1 Flash 1017 (max) | 47% | 31% | 62% | 46 | — |
| Kimi Code CLIKimi K3 (max) | 43% | 21% | 60% | 41 | — |
| MirrorBeam 1.0 | 20% | 4% | 38% | 20 | — |
The three benchmarks expose the trade-off behind that average:
- Terminal-Bench 4.0: Fusion scored 71, the strongest result in this comparison on terminal-based tasks. This was its clearest win.
- FrontierCode v1.1: Fusion scored 51. Claude Code with Opus 5.5 Max scored 56, while Devin Fusion with Opus 5.5 Max and GPT-6 Luna at High scored 54. Fusion trails here on repository coding tasks.
- SWE-Atlas-QnA-AA: Fusion scored 57 on questions that require understanding a code repository. Artificial Analysis characterizes this as a relative weakness compared with other agents in the table.
So the aggregate score of 60 masks different profiles: Fusion led Terminal-Bench but not the other two tests. The index design matters too: each component contributes one third. The comparison is between complete systems with their agents and declared configurations. Results in your own repository can change with the task, tools, and organization policies.
The practical reading is narrow: Fusion came close to this Claude Code configuration at a lower observed API cost in this evaluation. The benchmark does not prove that orchestration alone caused the difference, or that every user will spend half as much. Cost and quality depend on the task, which models run, and how many steps they take.
How single, cascade, and critique work
GitHub documents three execution patterns. Single uses one model to complete the task. Cascade asks an efficient model for a first draft, then a judge decides whether it is good enough or should be escalated to a stronger model. Critique starts with a draft, gathers independent critiques in parallel, and asks the original model to revise its answer using that feedback. Different prompts in one conversation can use different patterns. A simple task should not wait for a review it does not need; Cascade and Critique can take longer and consume more credits because they add model calls. (official documentation)
Standard, High, and Max describe levels of effort, from everyday work to more demanding reasoning. They do not promise a fixed set of models. The pool changes as GitHub adds models and evaluates their performance, and organization policies can exclude models. Administrators can control which models Fusion may use and whether it is the default.
How work gets routed
Solve with one model
One model completes the request. A straightforward task avoids extra review calls.
A judge decides whether to escalate
An efficient model drafts. A judge accepts the draft or escalates to a more capable model.
Parallel critiques, then a revision
One model drafts; several models critique independently. The original model revises its answer using their feedback.
An explanation of documented patterns, not a live run. The execution pattern and models can change with each request. GitHub documentation ↗
Who can use Copilot Fusion, and how is it billed?
GitHub says Fusion is available across Copilot plans in Visual Studio Code, Copilot CLI, the GitHub Copilot app, and JetBrains IDEs. Access to each effort level depends on the plan, enabled models, and administrator policies; GitHub does not give one subscription price for Fusion in this announcement. To try it, update the client, open the model picker, select Copilot Fusion, and choose an effort level. Fusion may not appear if the account cannot access any model in its pool. Official plans and policies.
Billing is based on each model used at its standard rate; Fusion has no separate fee. High and Max can call more than one model and use more AI credits than Standard. GitHub is offering users on paid plans a temporary 30% discount on Fusion model costs through November 30, 2026. The $5.59 and $11.42 figures are API costs per task reported by Artificial Analysis. We do not apply another 30% to those numbers: the evaluation does not explain the promotion’s treatment here. Check the current rates and AI credit terms for your account.
The practical limitation is central to the product’s promise: routing hides model selection, but it does not remove the need to review the result. GitHub’s documentation says Fusion works separately from subagents and does not start them. If Fusion discards a draft that has already changed the workspace, those edits are not automatically undone. Its execution patterns and model pool can also change. Fusion reduces the manual work of choosing which model to try; developers still check the result and decide what gets merged.
GitHub’s original announcement
GitHub introduced Copilot Fusion on October 28, 2026. The original post compares Fusion Max with Claude Code and links to access and promotion terms.
Source
Plan your next post in HeyMark.
Keep the idea, review the draft with your team, and see how it performed in the accounts you connected.