GitHub Copilot Fusion: not a model, but a director choosing between Anthropic and OpenAI

GitHub announced on October 28 in its changelog, during its Universe conference, that Copilot Fusion is now available to everyone. In the model picker it looks like any other entry, but it isn't. For each prompt it chooses which models to use and how to combine them. The user only sets the effort level: Standard, High or Max.

According to GitHub, there are three ways of working. In Single, one model handles the task. In Cascade, a cheap model writes a draft and a quality check either approves it or passes it to a more powerful model. In Critique, one model writes, a second reviews, and the first revises. GitHub's chief product officer, quoted by Let's Data Science, summed it up this way: “It's not a model. It's an orchestration engine at runtime.”

Fusion merges Project HydraFusion, the research preview launched in September, with the automatic model selection Copilot already offered. Anyone who was using auto, HydraFusion or HydraFusion Max is moved automatically to the matching Fusion level.

The changelog doesn't say which models are in the pool. According to Artificial Analysis, they include Claude Opus 5.5, Claude Sonnet 5.5, GPT-6 Astra and GPT-6 Sol — models from both Anthropic and OpenAI. The firm tested Fusion at the Max level in GitHub Copilot CLI using its Coding Agent Index v2.0. The result is 2 points below Claude Code with Opus 5.5 (max). GitHub gives Fusion Max a score of 60, but that absolute figure doesn't appear in the Artificial Analysis article. When we checked the index's interactive page, Fusion wasn't listed. In the same analysis, Devin Fusion, which combines Opus 5.5 (max) and GPT-6 Luna (high), scores 62. In the individual tests, Fusion scores 71 on Terminal-Bench 4.0, the highest value on the chart. On FrontierCode it stops at 51 and comes third, behind Claude Code with Opus 5.5 (max) at 56 and Devin Fusion at 54. On SWE-Atlas-QnA-AA it scores 57.

The real argument is cost. Artificial Analysis estimates an average API cost per task of $5.59 for Copilot Fusion, against $11.42 for Claude Code with Opus 5.5 (max) and $8.00 for Devin Fusion. Fusion uses about 7.2 million tokens per task: roughly half as many as Claude Code with Opus 5.5 and 30% fewer than Devin Fusion. GitHub adds a figure on time, about 40% less per task than Claude Code with Opus 5.5 Max, which Artificial Analysis doesn't report. GitHub also quotes a principal software engineer at AT&T who says Fusion Standard improved token efficiency at the company. But the $5.59 measures the benchmark's API cost, not what a subscriber spends. Monthly fees, failed tasks and human review time are left out. For users, every model Fusion calls is billed at its standard rate. The changelog announces a 30% discount on models used through Fusion for paid plans, valid for a limited period of unspecified length.

Fusion is available on all Copilot plans: in Visual Studio Code, the GitHub Copilot app, Copilot CLI and JetBrains IDEs. Support for other clients is described as coming soon. Enterprise admins can see which models Fusion chooses from, exclude a model from Fusion only, and set Fusion as the default model. The changelog is explicit: Fusion only uses models enabled in enterprise and organization policies. It follows that the actual pool varies from one organization to another and from one request to the next, and the full list isn't public. According to Let's Data Science, choosing between local and cloud models is experimental and must be switched on by the user. The managed local model, MAI Code 1.1 Flash, can only be downloaded on Windows. Even with a local model the session doesn't work offline, because the decision about which model to use is made in the cloud. It's a small detail that says a lot: the model can run on your machine, but the choice of which model to use still happens in the cloud.

What strikes me is that the most interesting product this autumn isn't a new model but a way of routing work among the ones that already exist. If the numbers hold up, owning the strongest model will matter less and less, and deciding which one to call, and when, will matter more and more. To judge it, though, we'd need data we don't have yet: an absolute score verified by third parties, independently measured times, and a bill that includes what the benchmark leaves out. Until then, I'll take Fusion for what's documented: nearly on par with the best, at about half the API cost of Claude Code with Opus 5.5.

— Olya