GitHub Copilot Fusion Orchestrates Anthropic and OpenAI Models for Lower-Cost Coding Tasks

GitHub Copilot Fusion Orchestrates Anthropic and OpenAI Models for Lower-Cost Coding Tasks

GitHub has introduced Copilot Fusion in GitHub Copilot, a runtime orchestration system that selects and coordinates models for each task. The system is designed to balance performance, cost, and latency across models from Anthropic and OpenAI.

Copilot Fusion can choose among models including Claude Opus 5.5, Claude Sonnet 5.5, GPT-6 Astra, and GPT-6 Sol. Its maximum setting in GitHub Copilot CLI was evaluated in the Artificial Analysis Coding Agent Index v2.0. Token charges are based on the underlying models used to complete each request.

Benchmark and cost results

The Coding Agent Index v2.0 gives equal weight to its three constituent benchmarks. Copilot Fusion achieved the following results:

• Cost per task: $5.59, compared with $11.42 for Claude Code using Opus 5.5 at its maximum setting.
• Terminal-Bench 4.0: 71, the highest score among the systems described in the comparison.
• FrontierCode: 51, placing it third in the chart behind Claude Code with Opus 5.5 at maximum, which scored 56, and Devin Fusion with Opus 5.5 at maximum plus GPT-6 Luna at high, which scored 54.
• SWE-Atlas-QnA-AA: 57, between GPT-6 Astra in Codex and Gemini 4 Argon in Antigravity CLI.

Devin Fusion with Opus 5.5 at maximum and GPT-6 Luna at high scored 62 overall at a cost of $8.00 per task.

Token usage

Copilot Fusion used 7.2 million tokens per task in the Coding Agent Index. That is approximately half the usage recorded for Claude Opus 5.5 in Claude Code and about 30% less than Devin Fusion with Opus 5.5 at maximum plus GPT-6 Luna at high.

The results place Copilot Fusion near the leading systems in the evaluated coding benchmarks while using fewer tokens per task than the cited alternatives.