GitHub Copilot Fusion nears Opus 5.5 at half the benchmark cost

GitHub used its October 28 Universe conference to introduce Copilot Fusion, a system that routes coding requests to one or more models and coordinates their work. In an Artificial Analysis benchmark published the same day, Fusion Max averaged $5.59 per task, versus $11.42 for Claude Code with Claude Opus 5.5 in Max mode. Fusion scored 60 on the Index coding-agent benchmark, two points behind Claude Code. The result gives GitHub independent evidence for its pitch: choosing and coordinating models can make AI-assisted coding cheaper without giving up much benchmark performance.

A coordinator, not another model

Fusion assesses tasks involving reasoning, code generation, debugging and tool use, then decides whether to send a request to one model or coordinate several. It combines capabilities developed for GitHub’s Project HydraFusion research effort with automatic model selection.

Developers now choose an effort level rather than a separate Fusion model:

Fusion can handle a task in three ways:

Artificial Analysis found models from Anthropic and OpenAI in Fusion’s model pool: Claude Opus 5.5, Claude Sonnet 5.5, GPT-6 Astra and GPT-6 Sol. The model selected depends on the request and the organization’s policies.

That flexibility is the product’s point: developers describe what they want the software to do, while Fusion decides how to get there. But the benchmark does not show whether the same approach reduces the total work of delivering software.

What the benchmark measured

Artificial Analysis tested Fusion Max through GitHub Copilot CLI on Index v2.0, a coding-agent benchmark that gives equal weight to three programming evaluations. Fusion scored 60; Claude Code with Opus 5.5 Max scored about 62. Devin Fusion, using Opus 5.5 Max and GPT-6 Luna High, also scored 62.

Fusion used 7.2 million tokens per task, roughly half the amount used by Claude Code with Opus 5.5 and about 30% less than Devin Fusion. That suggests its cost advantage comes at least partly from doing less model work, not only from using cheaper tokens.

Scores varied by task. Fusion led the systems covered in the report on Terminal-Bench 4.0, with 71 points. On FrontierCode, which tests changes to repositories, it scored 51, behind Claude Code with Opus 5.5 at 56 and Devin Fusion at 54. On SWE-Atlas-QnA-AA, a repository question-and-answer test, it scored 57. Its near-tie on the overall index does not mean Fusion wins every kind of coding task.

GitHub CEO Satya Nadella has cited potential savings of up to 77%, but that figure comes from a separate company evaluation, not the Artificial Analysis average. GitHub’s internal tests used DeepSWE and its Copilot agent system; Fusion High and Max beat comparable frontier-model configurations by as much as 8.7 points on task completion, with estimated costs falling by up to 77%. Those conditions differ from the independent comparison, and the two savings figures are not directly comparable.

The independent report also does not include task-by-task completion times to verify GitHub’s claim that Fusion Max was about 40% faster than the Claude Code configuration. Its cost figures describe benchmark tasks, not what a specific company would pay under GitHub’s new AI-credit purchasing option.

The gap between a cheaper benchmark run and cheaper software delivery is where the harder questions start. Teams still have to account for human review, errors, retries, integration and security work after an agent responds. I think the token numbers make Fusion’s efficiency case more concrete, but they do not yet settle the business case.

The price of routing

GitHub says Fusion is available to all Copilot subscribers in VS Code, the Copilot app, Copilot CLI and JetBrains development environments, with support for other clients planned. Users pay the standard rate for each model Fusion calls. Paid Copilot subscribers temporarily receive a 30% discount on model costs when using Fusion; that promotion is separate from the benchmark’s measured savings.

GitHub says routing and coordination, rather than subsidizing expensive requests, make Fusion economical. That does not disclose GitHub’s margins or guarantee that every task will cost less than choosing a model manually.

Separately, some GitHub Enterprise Cloud customers can buy Copilot through an enterprise budget, without purchasing an additional Copilot subscription for every licensed Enterprise user. Under the published terms, there is no monthly AI-credit allowance per employee: usage-based charges begin with the first request.

Code completions and next-edit suggestions remain available if the AI budget runs out. Copilot Chat, code review and Copilot in the cloud require an available budget. Copilot Business and Copilot Enterprise remain on sale, while self-service access to the new purchasing option is being phased in. GitHub also lists Copilot Standalone at $10 per user per month for enterprise-account users without a GitHub Enterprise license.

The terms explain who can use the new option and how charges work, but do not show what a typical development team would spend in a month. Buyers will have to weigh actual usage and budget limits against plans priced by seat.

The rest of the workflow

GitHub is extending Copilot beyond code suggestions. Copilot Projects lets teams run parallel cloud-agent sessions across repositories and track them in one interface. Session continuity lets developers move between the Copilot app, website, command line and VS Code while keeping the same cloud agent’s context. GitHub says work can continue across devices, including mobile.

Copilot Automations will let teams run agent workflows manually, on a schedule or in response to events. Teams can review and approve suggested improvements based on earlier results; approval remains with the human.

Fusion also raises a governance question: can companies restrict which models it selects? GitHub says Fusion follows enterprise and organization model policies. Administrators can see the models available to Fusion, disable individual models for it and set a default effort level. In an interview with VentureBeat, GitHub chief product officer Mario Rodriguez confirmed that users can see which models are being used, but did not specify what details will be available to audit each request.

As AI-generated code grows, GitHub is also adding a maximum code-review level, separate from Fusion Max, alongside Lite and Balanced. The company says the new level has the best F1 score for serious issues on ReviewBench, but its Universe announcement included no detailed results breakdown.

GitHub is expanding the infrastructure behind these products, too. The company said it added 160 petabytes of high-speed storage, 4 million CPU cores and network capacity of 50 terabits per second in 2026. It plans to complete a move to Azure deployments across multiple regions in early 2027 and is changing Git’s architecture to separate compute from storage.

Rodriguez linked that work to more agent-driven reads and writes in repositories. More parallel sessions fetching code, changing files and submitting updates put added demands on development infrastructure.

The meaningful test for Fusion is therefore not whether it can approach Opus 5.5’s benchmark score at roughly half the measured cost. It is whether that advantage survives code review, access controls, reliability and the full release cycle—and leaves teams with more quality software for less.