By Paul Derrington, Jr.
CEO | Chief Engineer | Research Engineer | Editor, D-CSIL
3 min read
Key Takeaways
GitHub made Copilot Fusion generally available on October 28, at GitHub Universe. Instead of picking a model for each coding task, you pick Fusion and an effort level, and Copilot routes the work across models, using one model, a draft-and-check cascade, or a draft-and-critique pass.
What changed on October 28?
Copilot Fusion is generally available, GitHub said in a changelog post on October 28, announced at GitHub Universe in San Francisco. It replaces the job of picking a model in the model picker: select Copilot Fusion, choose an effort level, and Copilot routes and orchestrates the models for each prompt.
Fusion combines two earlier pieces: Project HydraFusion, the research preview for multi-model orchestration, and Copilot auto model selection. People who had enabled auto, HydraFusion, or HydraFusion Max move to the matching Fusion effort level automatically. As Jay Parikh, Microsoft's executive vice president of CoreAI, whose teams include GitHub, put it, quoted by The Register: “Developers shouldn’t have to manage every model choice to get the right result at the right price.”
How does Fusion decide which model does the work?
GitHub says that for each prompt, Fusion weighs what the task needs across reasoning, code generation, debugging, and tool use, then picks how to run it. You choose the effort: Standard for everyday coding with versatile models, High for harder tasks and deeper reasoning, and Max for the most demanding work with the most capable models.
Fusion runs one of three patterns. Single: one model solves the task directly. Cascade: an efficient model drafts, a quality check accepts it or escalates to a stronger model. Critique: one model drafts, a second model reviews it, and the first model revises. GitHub says you can watch Fusion analyze the task and pick; the routing itself is not something you configure prompt by prompt.
What do the benchmark figures show?
GitHub says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code running Opus 5.5 Max, at about half the average API cost per task, about half as many tokens, and about 40 percent less time per task. Artificial Analysis, which runs the index, published its own results the same day: it tested Fusion's Max setting in GitHub Copilot CLI on version 2.0 of the index and reported a score of 60 against 62 for Claude Code with Opus 5.5 (max), $5.59 per task against $11.42, and 7.2 million tokens per task, about half of Claude Code's. GitHub's 40 percent time figure is not in Artificial Analysis's write-up. The results were uneven across the index's three equally weighted tests: Artificial Analysis lists Fusion first on Terminal-Bench 4.0 at 71, third on FrontierCode at 51, behind Claude Code with Opus 5.5 at 56, and at 57 on SWE-Atlas-QnA-AA. These are benchmark API costs, not a forecast of any team's bill.
The one outside voice in GitHub's post is AT&T principal software engineer Hari Pokhrel, who said Fusion Standard improved token efficiency across AT&T by picking the right model per task, cutting unnecessary spend. He also said developers used to pick a powerful model for a hard task and forget to switch back for routine work. That is a single customer's experience, not a benchmark.
What does it cost, and who stays in control?
You are billed for each model Fusion uses, at each model's standard rate. For a limited time, paid Copilot plans get a 30 percent discount on model costs when using Fusion; the changelog gives no end date, and GitHub's announcement post on X says the discount runs through November 30. Fusion is available on all Copilot plans in Visual Studio Code, the GitHub Copilot app, the Copilot CLI, and JetBrains IDEs, with more clients coming.
Administrators keep the guardrails. Fusion only uses models enabled in the enterprise and organization model policies, admins can disable a model for Fusion without disabling it for everyone, and they can set Fusion as the default. Because Fusion chooses the models, it also determines which model rates apply to each prompt, so teams can compare their own cost per task with the benchmark averages.