GitHub wants developers to stop choosing models. At its Universe conference on October 28, the company made Copilot Fusion generally available across all Copilot plans. Instead of selecting a specific model from Copilot's picker, a user selects Fusion and an effort level, and the system decides for every prompt which models to call and whether to combine them.
Fusion merges two earlier pieces of GitHub's product: the multi-model orchestration tested in the Project HydraFusion research preview and the model routing behind Copilot's automatic selection. Users pick Standard for routine coding, High for harder problems or Max for the most demanding work. Behind each request GitHub uses one of three patterns. In Single mode one model does the job. In Cascade, a cheaper model drafts and a quality check either accepts the draft or hands it to a stronger model. In Critique, one model answers, a second reviews and the first revises.
Cost is the pitch. GitHub says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code running Claude Opus 5.5 Max, at about half the average API cost and half the tokens per task, and in about 40% less time. VentureBeat, which reviewed the Artificial Analysis results, reports an average of $5.59 per task for Fusion Max against $11.42 for the Claude Code configuration and $8.00 for Cognition's Devin Fusion, which scored 62. The models Fusion drew on during the test came from Anthropic and OpenAI; Artificial Analysis named Claude Opus 5.5 and Claude Sonnet 5.5 on one side, GPT-6 Astra and GPT-6 Sol on the other.
The detailed numbers are less uniform than the headline. According to VentureBeat, Fusion led the group on Terminal-Bench 4.0 with 71 but trailed on FrontierCode, a test of repository changes, with 51 against 56 for Claude Code and 54 for Devin Fusion. The 40% time saving does not appear in the independent data and remains GitHub's own figure, as does a 77% saving cited by Microsoft chief executive Satya Nadella, which refers to a separate internal evaluation.
Enterprises keep control over which providers are used. Fusion only draws on models enabled in an organization's policies, administrators can exclude a model from Fusion without disabling it elsewhere, and they can make Fusion the default at any effort level. Billing follows each underlying model's standard rate, with a 30% discount on paid plans for a limited period. It is available in Visual Studio Code, Copilot CLI, JetBrains IDEs and GitHub's Copilot app, and in JetBrains it replaces the previous Auto option. GitHub chief product officer Mario Rodriguez told VentureBeat that the system is an orchestration engine running at request time, not a model.
A second release the same day extends routing to the device. In an opt-in experimental preview, Fusion Standard can choose between cloud models and a supported local model, such as MAI Code 1.1 Flash installed through Windows ML, or models like Qwen3.8 Flash Next served from a local OpenAI-compatible endpoint in Copilot CLI. GitHub notes that routing decisions still happen in the cloud and that local models may carry different safety protections.
Fusion turns model choice into a managed service, which could cut spending for teams that habitually run every task on the most expensive model. Whether benchmark savings translate into cheaper accepted code, once review time and failures are counted, is something customers will have to measure for themselves.