GitHub Copilot Fusion Max scored 60 on an independent coding-agent index. Claude Code with Opus 5.5 scored 62. Fusion finished each test task for $5.59; Claude Code needed $11.42.
That is half the price for a gap of two points, and it matters most to developers who pay per model call and run long agent tasks. It is a cost per benchmark task, not your monthly Copilot bill.
The short answer
- Score: Fusion Max in Copilot CLI scored 60 on the Artificial Analysis Coding Agent Index v2.0. Claude Code with Opus 5.5 (Max) scored 62.
- Cost per task: $5.59 for Fusion, $11.42 for Claude Code. Fusion also used about 7.2M tokens per task, roughly half as many.
- What Fusion is: a model picker option that routes each prompt across several models. It is not a new model.
- Billing: you pay each model’s standard rate, with no extra Fusion fee. Paid plans get 30% off model costs through November 30, 2026.
- Plans: Standard works on every plan. High needs Pro+ or higher. Max needs the Max, Business or Enterprise plan.
What is Copilot Fusion, and how does it choose a model?
Fusion became generally available on October 28, 2026. You pick Copilot Fusion in the model picker, then pick an effort level: Standard, High or Max. Fusion decides the rest, one prompt at a time, and one session can mix all three patterns.
- Single: one model answers.
- Cascade: a cheaper model drafts, and a quality check accepts the draft or sends the task to a stronger model.
- Critique: one model drafts, other models review it, and the first model revises.
In the Artificial Analysis test, Fusion routed across Claude Opus 5.5, Claude Sonnet 5.5, GPT-6 Astra and GPT-6 Sol. The design explains the savings: Cascade lets an efficient model draft first, and a stronger model steps in only when the quality check rejects the draft.
How did Fusion compare with Claude Code?
| System tested | Coding Agent Index v2.0 | Cost per task |
|---|---|---|
| Copilot Fusion Max in Copilot CLI | 60 | $5.59 |
| Claude Code with Opus 5.5 (Max) | 62 | $11.42 |
| Devin Fusion with Opus 5.5 (Max) and GPT-6 Luna (high) | 62 | $8.00 |
Read the table this way. Fusion gives up about 3% of the score (2 of 62) and saves about 51% of the cost. Devin Fusion matches Claude Code at 62 and costs about 30% less. If your bottleneck is cost, both Fusion systems beat Claude Code in this test. If your bottleneck is the last two points of quality, Claude Code still leads.
GitHub’s own announcement adds that Fusion Max takes about 40% less time per task. Treat that as GitHub’s claim; the independent article covers score, cost and tokens. GitHub’s announcement is also the source of the score of 60, and Artificial Analysis confirms the two-point gap.
Where does Fusion win, and where does it trail?
The index weighs three benchmarks equally, so a single strong result can hide a weak one.
- Terminal-Bench 4.0: Fusion scores 71, the top result in the chart.
- FrontierCode: Fusion scores 51 and ranks third. Claude Code scores 56 and Devin Fusion 54.
- SWE-Atlas-QnA-AA: Fusion scores 57.
The pattern is clear. Fusion is strong on terminal-style tasks and weaker on FrontierCode, where Claude Code leads by five points. If your daily work is command-line automation, Fusion’s lead is relevant. If it is hard, open-ended coding problems, the FrontierCode gap is the number to watch.
Why is cost per task not your monthly bill?
Artificial Analysis charges each request at the API rate of the model that handled it, then adds up the cost of finishing the benchmark tasks. That number ignores your subscription fee, your plan’s credits and how many tasks you run.
It also tests an agent and its models together, so you cannot credit the gap to the models alone. And do not subtract the 30% discount from these figures. The source does not say whether it was applied, so the discount may already be inside them or not at all.
On your own bill, expect this: High and Max can use several models for one task, so a task may burn more AI credits than on Standard. Fusion tries to keep the main conversation on one model, which preserves cached tokens, and gives helper models such as reviewers only the context they need.
Which plan do you need?
| Plan | Standard | High | Max |
|---|---|---|---|
| Free, Student, Pro | Yes | No | No |
| Pro+ | Yes | Yes | No |
| Max, Business, Enterprise | Yes | Yes | Yes |
Fusion runs in VS Code, the GitHub Copilot app, Copilot CLI and JetBrains IDEs. It picks only from models your plan and your organization’s policies allow, including data-residency and FedRAMP limits. So your model pool may differ from the test pool. If no allowed model exists, Fusion does not appear in the picker.
Enterprise admins can see which models Fusion chooses from, disable specific models for Fusion only, and set Fusion at any effort level as the default model.
Can Fusion run local models?
Yes, in an opt-in experimental preview, and only at Standard effort. MAI Code 1.1 Flash is a managed download on supported Windows devices. Qwen3.8 Flash Next works through a local server such as Ollama, on Windows, macOS and Linux. In the Copilot app, automatic routing works only with managed Windows models.
Local does not mean offline. Fusion still makes its routing decisions in the cloud. Local models may also carry different safety protections than cloud models.
What to do this week
- Find your plan in the table. If you are on Free, Student or Pro, you get Standard only, so the Max result does not apply to you.
- If you pay for Copilot, try Fusion before November 30, 2026, while the 30% discount lasts.
- Run the same three to five real tasks on Fusion and on your current tool. Compare time, cost and result. A benchmark average does not know your codebase.
- Hover over the effort level under a response to see which models Fusion used.
- Review file changes before you commit. If Fusion discards a draft that already edited files, those edits are not undone automatically.
Who should care: teams on Max, Business or Enterprise who run long agent tasks, and anyone watching AI credits. Who can ignore it for now: Free, Student and Pro users who only get Standard, and anyone who needs the highest FrontierCode-style quality.
Not announced yet: whether the 30% discount is inside the test figures, and what the discount will be after November 30, 2026.
Related: Gemini 4 Carbon: benchmarks, price and who can use it, which covers the Intelligence Index and what changed in v5.
Questions and answers
Is Fusion cheaper than Claude Code?
Per benchmark task, yes: $5.59 against $11.42, about half. Your subscription cost depends on your plan and credits, which the test did not measure.
Is Fusion a new AI model?
No. It is a router that picks existing models and a workflow for each prompt.
Does Fusion beat Claude Code at coding?
Not overall. It scored 60 against 62. It led on Terminal-Bench 4.0 (71) and trailed on FrontierCode (51 against 56).
Which plan do I need for Fusion Max?
The Max, Business or Enterprise plan. Pro+ gets Standard and High, and Free, Student and Pro get Standard.
Can I use Fusion offline?
No. Routing happens in the cloud, even when a local model does part of the work.
Does Fusion launch subagents?
No. GitHub’s docs say Fusion works separately from subagents and does not start them.
Terms in plain language
- Coding agent: a tool where a model reads and edits code and runs commands for you.
- Routing: choosing which model handles a task.
- Cost per task: what a benchmark run spent on one test task, at API prices.
- Token: a small piece of text models read and write.
Sources
- Artificial Analysis: Copilot Fusion results
- GitHub changelog: Copilot Fusion generally available
- GitHub Docs: Copilot Fusion
- GitHub changelog: local model selection
Nobody at ClueTrove ran these tools. This article reports published tests and documentation.
Facts read on 29 October 2026
