loading
loading
GitHub's Copilot Fusion Matches Opus Quality at Half the Cost GitHub's new runtime orchestration layer routes between Claude and GPT models, scoring near Opus 5.5 at half the cost per task. - GitHub Copilot Fusion is generally available, routing tasks across Claude and GPT-6 models. - Scores 60 on Artificial Analysis Coding Agent Index, two points behind Claude Opus 5.5 Max. - Costs $5.59 per task versus $11.42 for Claude Code with Opus 5.5 Max. - Uses 7.2M tokens per task, roughly half of Opus 5.5 in Claude Code. - Three orchestration modes: Single, Cascade (plan then execute), and Critique (debate). - Leads Terminal-Bench 4.0 at 71, but trails on FrontierCode at 51. GitHub has made Copilot Fusion generally available. Fusion routes coding tasks among Anthropic and OpenAI models, selecting a model and coordination strategy for each request. In the independent Coding Agent Index, Fusion Max scored 60 against Claude Code with Opus 5.5 Max at 62, while costing $5.59 per task compared with $11.42. Fusion takes over model choice Developers select Fusion and one of three effort levels: Standard, High, or Max. The available model pool includes Claude Opus 5.5, Claude Sonnet 5.5, GPT-6 Astra, and GPT-6 Sol. GitHub combines Project HydraFusion’s multi-model orchestration with the model selector previously used by Copilot’s automatic model mode. Copilot chooses among three coordination strategies for each request: - Single: One model handles the task from start to finish. - Cascade: One model creates the plan and another executes it. - Critique: One model proposes an approach, a second reviews it, and the models revise the plan before completing the task. According to GitHub, usage charges reflect the underlying models that Fusion invokes. Tokens are the pieces of text that models process, so lower token consumption usually reduces inference cost. Max cuts measured cost in half Artificial Analysis tested Fusion Max through the GitHub Copilot CLI. Its composite index weights Terminal-Bench 4.0, FrontierCode, and SWE-Atlas-QnA-AA equally, covering terminal work, difficult coding problems, and software-engineering questions. Costs and token counts come from benchmark runs and differ from Copilot subscription pricing. The approximate competitor token totals are derived from the reported ratios. Fusion used about half as many tokens as Claude Code and roughly 30% fewer than Devin Fusion, which combines Opus 5.5 Max with GPT-6 Luna. Those totals are consistent with a router assigning cheaper models to simpler subtasks, although the published figures do not isolate the effects of routing, shorter reasoning, or caching. Claude Code’s five-point advantage on FrontierCode shows where the aggregate score compresses a meaningful difference. The benchmark also covers one agent harness and three test suites, so repository structure, tool access, and task mix may produce different results in production. Two years of model expansion GitHub began moving Copilot toward a multi-model architecture at GitHub Universe in October 2024, when it added model choices from Anthropic, OpenAI, and Google. Fusion automates that selection and can divide one task among several models. GitHub’s customer account quotes an AT&T engineer who said developers often selected a powerful model for a difficult task and left it enabled for routine work. Fusion addresses that usage pattern automatically. Enterprise administrators can also set Fusion and an effort level as the default through managed settings. GitHub becomes the model broker Fusion places GitHub between model vendors that also sell competing coding agents. OpenAI offers Codex, while Anthropic offers Claude Code. Routing gives GitHub room to optimize cost and quality across suppliers without tying Copilot’s behavior to one model family. That abstraction introduces operational trade-offs. Cascade and critique strategies can require multiple sequential model calls, which may increase latency and make completion times less predictable. Latency data is absent from the cited benchmark results. Routed execution can also complicate reproducibility because the selected model and strategy may vary by task. Organizations with vendor restrictions should confirm which models their policies permit before enabling Fusion. The pool spans Anthropic and OpenAI, so reviews may need to cover model allowlists, data handling, regional requirements, and audit records. What to measure in production Teams using Copilot Business or Enterprise can evaluate Fusion Max with a controlled set of representative tasks. A useful comparison should track: - Task completion and code acceptance rates - Human rework after agent completion - Median and high-percentile latency - Token use and effective cost per completed task - Performance on routine edits, migrations, debugging, and unfamiliar code - Compliance with approved model and provider policies Developers operating their own agent harnesses can apply the same architecture: classify the request, assign a model, set a compute budget, and add planner or critic calls when the task warrants them. Implementing that pattern also requires reliable context transfer, retry handling, cost limits, and evaluation data. Production adoption should depend on measured results from the team’s repositories and workflows.
document fetched 2026-10-29 03:02