Overview
GitHub Copilot Fusion is a runtime orchestration system announced at GitHub Universe 2026 on October 28, 2026. It does not train models. It picks and coordinates them per task, drawing from Claude Opus 5.5, Claude Sonnet 5.5, GPT-6 Astra and GPT-6 Sol, and solves for performance, cost and latency at the same time.
Artificial Analysis benchmarked its max setting in GitHub Copilot CLI against the Coding Agent Index v2.0. Cost per task is 5.59 dollars, versus 11.42 dollars for Claude Code with Opus 5.5 at max. The aggregate score lands two points below that configuration. Token use is 7.2 million per task, about half of Opus 5.5 in Claude Code and 30 percent less than Devin Fusion with Opus 5.5 plus GPT-6 Luna.
On the three constituent benchmarks it leads Terminal-Bench 4.0 at 71, ranks third on FrontierCode at 51 and scores 57 on SWE-Atlas-QnA-AA. Token costs are billed against whichever underlying model handled the request; the orchestration layer itself adds no surcharge.
Key Features
- Cross-vendor model selection: The pool spans Anthropic and OpenAI. A single task can be decomposed by one model and executed by another. Selection follows task characteristics rather than a fixed routing table.
- Half the cost: 5.59 dollars per task against 11.42 dollars for Claude Code with Opus 5.5 at max in the same index. Token use of 7.2 million per task is roughly half.
- Leads Terminal-Bench 4.0: Scores 71 on that benchmark, first among the compared agents. FrontierCode comes in at 51 for third, and SWE-Atlas-QnA-AA at 57.
- Harness open to developers: Microsoft concurrently opened the Copilot harness built by the GitHub team through Microsoft Foundry. Copilot Code, Copilot Studio, the Copilot app and the CLI all run on it.
- Delivery chain rebuilt in the same release: Universe 2026 also brought the Copilot app, Copilot projects for parallel cross-repository sessions, session continuity across app, VS Code, CLI and web, and code review Max. Enterprise billing moved to AI credits.
- Billed by underlying model: The orchestration layer does not add a margin; invoices follow the tokens of whichever model ran. That keeps the markup transparent and makes cost predictability depend on usage.
Use Cases
- Cost-sensitive coding work that still needs near-frontier quality.
- Multi-repository projects run in parallel through Copilot projects.
- Heavy terminal, script and build work, where the Terminal-Bench 4.0 result maps directly.
- Enterprises standardising on one delivery chain across app, VS Code, CLI and web.
- Review stages that need better recall on high-severity issues through code review Max.
- Teams embedding the same harness in their own products via Microsoft Foundry.
Pros
- 5.59 dollars per task, about half of the comparable Opus 5.5 setup.
- Aggregate score only two points behind Opus 5.5.
- Leads Terminal-Bench 4.0 at 71.
- 7.2 million tokens per task, 30 percent less than Devin Fusion with Opus 5.5 plus GPT-6 Luna.
- No surcharge on orchestration; billing follows the models actually used.
- Harness available through Microsoft Foundry for in-house products.
Pricing
Used within a GitHub Copilot subscription. Token costs are billed against the underlying model that handled each request, with no extra charge from the orchestration layer. Enterprise plans moved to AI credits billing after Universe 2026; consult GitHub's pricing page for credit rates and allowances. GitHub has not published a separate Fusion subscription tier.
Summary
Fusion addresses a concrete problem: frontier models are expensive per task, and most tasks do not need the top tier. Moving model selection into the runtime trades a fixed routing policy for cost. The Artificial Analysis numbers show the trade works, two points of score for half the money and less time. The cost is predictability, since the bill follows the underlying models, so budget in tokens rather than in runs.
Version History
- GitHub Copilot Fusion released: orchestration across Anthropic and OpenAI (2026-10-29): GitHub Universe 2026 introduced Copilot Fusion, a runtime that selects among Claude Opus 5.5, Sonnet 5.5, GPT-6 Astra and GPT-6 Sol per task. Artificial Analysis Coding Agent Index v2.0 measured 5.59 dollars per task against 11.42 dollars for Opus 5.5 in Claude Code, scoring two points lower overall; Terminal-Bench 4.0 at 71 for first, FrontierCode 51 for third, SWE-Atlas-QnA-AA 57; 7.2 million tokens per task, about half of Opus 5.5. The same event brought the Copilot app, Copilot projects, session continuity, code review Max and AI credits billing for enterprise.