Microsoft subsidiary and popular code repository hosting service GitHub is rolling out a system that decides which AI models should handle a software development task — and, according to independent testing, can approach the performance of a leading coding agent for roughly half the expense.
At GitHub Universe conference on October 28, the company unveiled the new GitHub Copilot Fusion, a system that routes requests among models and coordinates them within a task.
In testing published October 28 by Artificial Analysis, Fusion Max cost an average $5.59 per task, compared with $11.42 for Claude Code running Anthropic's Claude Opus 5.5 at its Max setting — approximately 51% less — while finishing just two points behind on the firm's Coding Agent Index.
The results give GitHub an outside measurement to support its argument that intelligent model selection can make expensive AI-assisted development more economical.
GitHub also says Fusion Max completed those benchmark tasks in about 40% less time than the Claude Code configuration.
Microsoft CEO Satya Nadella has promoted potential savings of up to 77%, but that higher figure refers to a separate, company-reported evaluation, not the average measured by Artificial Analysis. Together, the claims put task-level costs, rather than the price of any one foundation model, at the center of GitHub's pitch to software teams.
In an interview with VentureBeat, GitHub Chief Product Officer Mario Rodriguez said Fusion is intended to make model selection less of a burden while improving the economics of completed work. The distinction matters for enterprise customers evaluating whether lower AI inference costs translate into less overall engineering effort.
Fusion is an orchestration engine, not another AI model
Fusion evaluates work 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 in GitHub's earlier Project HydraFusion research preview with its automatic model-selection feature.
“It's not a model. It's an orchestration engine at runtime,” Rodriguez told VentureBeat. He said Fusion can both route individual requests and orchestrate multiple model calls in pursuit of completing a task.
Rodriguez said GitHub has also changed how developers encounter the technology. Rather than treating it as another individual model in the lineup, GitHub now presents it as a selectable mode with different effort levels. “You just enable it, and then you have three reasoning efforts: standard, high and max,” he said.
Standard prioritizes versatile models for everyday coding
High targets more difficult work
Max aims at the most complex tasks.
The system uses three execution patterns:
Single sends the work directly to one model
Cascade begins with a less costly model and checks whether to accept its output or escalate it to a stronger one
Critique has one model draft an answer, another review it, and the original model revise its response. Those distinctions are outlined in GitHub's launch documentation.
“We have a single mode where we just go directly into the model that we believe is the best for the task,” Rodriguez explained. He also described Critique as a way to bring more capable intelligence into the review process rather than assign the most expensive model to every stage. The purpose, he said, is to let people focus on what they want their software to accomplish: “They don't have to be constantly thinking about a model.”
Artificial Analysis identified models from both Anthropic and OpenAI in Fusion's pool, including Claude Opus 5.5, Claude Sonnet 5.5, GPT-6 Astra and GPT-6 Sol. The specific models used can vary by request and by the policies an organization sets.
What the independent benchmarks show — and what they don't
Artificial Analysis tested Fusion's Max setting through GitHub Copilot CLI on its Coding Agent Index v2.0, which weights three coding evaluations equally.
Fusion earned an overall score of 60, compared with approximately 62 for Claude Code with Opus 5.5 Max. Competing orchestration system Devin Fusion, using Opus 5.5 Max and GPT-6 Luna High, scored 62 with an average task cost of $8.00.
On cost, Fusion's $5.59 average compared favorably with $11.42 for Claude Code with Opus 5.5 and $8.00 for Devin Fusion. Artificial Analysis also measured 7.2 million tokens per task for Fusion, approximately half the token use of the Opus 5.5 Claude Code configuration and about 30% less than Devin Fusion. That suggests the savings came at least in part from doing less model work, not merely using cheaper tokens.
Performance was uneven across the index's components, however. Fusion led the systems discussed in the report on Terminal-Bench 4.0 with a score of 71. On FrontierCode, a coding evaluation focused on repository changes, it scored 51, behind Claude Code with Opus 5.5 at 56 and Devin Fusion at 54. Fusion scored 57 on the SWE-Atlas-QnA-AA repository question-answering evaluation. In other words, the close composite score does not mean Fusion wins every kind of coding assignment.
Those are benchmark costs, not a published estimate of how much a particular enterprise will pay under GitHub's new AI-credit purchasing arrangement. The Artificial Analysis article also does not report the per-task completion-time data underlying GitHub's claim of 40% less time; that remains a GitHub-reported comparison, even though the performance, token-use and cost measurements have independent support.
The larger 77% figure needs further separation. As reported by Ken Yeung in The AI Economy, GitHub's own tests with the DeepSWE benchmark and Copilot's agent harness showed Fusion High and Max beating comparable frontier-model configurations by up to 8.7 points on task completion, while reducing estimated costs by as much as 77%.
The comparison involves different test conditions and a maximum reported reduction; it should not be conflated with Artificial Analysis's approximately 51% average cost advantage against Opus 5.5. The underlying DeepSWE configurations and per-task results were not available in the materials reviewed for this article.
GitHub's business case rests on an important distinction: saving money on a benchmark run is not necessarily the same as reducing the total cost of shipping software. Teams still need to account for human review, mistakes, retries, integration and any security work that follows an agent's output. Jay Parikh, Microsoft's executive vice president of CoreAI, emphasized the broader target in GitHub's Universe announcement: “Completed work, not more code, is the output that matters.”
Model-level billing, a temporary discount and a new enterprise option
Copilot Fusion is generally available across GitHub's Copilot plans in VS Code, the Copilot app, Copilot CLI and JetBrains IDEs, with support for additional clients planned. GitHub says users are billed for each underlying model Fusion calls at that model's standard rate.
It is also offering paid Copilot subscribers a temporary 30% discount on model costs when using Fusion. That promotional reduction is separate from the task-cost savings measured in benchmarks.
Rodriguez attributed Fusion's economics to its routing and orchestration choices rather than to GitHub subsidizing expensive requests. His explanation does not establish GitHub's profit margin or guarantee that every workload will be cheaper than simply choosing one model manually.
Separately, GitHub is giving eligible GitHub Enterprise Cloud customers another way to purchase Copilot. Instead of buying an additional Copilot subscription for every licensed Enterprise user, companies working through GitHub or Microsoft account teams can have access included and pay for AI usage against an enterprise budget.
The published terms specify that these users do not receive a monthly AI-credit allowance per seat; metered usage begins with the first request.
The distinction is not academic. Code completion and next-edit suggestions remain available even when no AI budget remains, but Copilot Chat, code review and cloud-based Copilot require available budget.
Existing Copilot Business and Copilot Enterprise subscriptions are still offered, and self-service access to the new arrangement will arrive in stages. GitHub also lists Copilot Standalone at $10 per user per month for enterprise-account users without a GitHub Enterprise license.
These details clarify access and billing structure, but the materials reviewed still do not provide a worked enterprise example showing the monthly AI-credit consumption of a typical development team under the new arrangement. Buyers will need to compare their actual usage and budget limits with existing seat-based terms.
Projects, persistent sessions and a wider audience
GitHub is extending Copilot beyond code suggestions into a place where developers can plan work, delegate assignments and inspect results. The Copilot app supports specifications, design discussions and telemetry analysis alongside programming.
Copilot Projects lets users direct parallel cloud-based agent sessions across several repositories and keep track of them in one interface.
Session continuity allows developers to move between the Copilot app, website, command line and VS Code without discarding the same cloud agent's context.
Rodriguez confirmed support for continuing work across devices, including mobile. That could matter to teams whose agents continue working after a developer closes a laptop or moves to another part of a project.
Rodriguez also framed the changes as a way to make development accessible to nonprofessionals without limiting experienced engineers. “This is for everyone,” he said. He described GitHub's ambition as lowering the barrier to creating software while giving skilled developers more leverage, and pointed to the platform's community of more than 225 million developers.
He also discussed GitHub's effort to retain useful information from agent sessions and feed it into organizational knowledge about technical domains, workflows and development practices.
GitHub is expanding Copilot Automations as well: teams can start agent workflows manually, schedule them or trigger them with events, then review and approve improvements the automations suggest based on past results. That approval requirement preserves a human decision point when a workflow proposes changing its own instructions.
Enterprise model controls and code quality
Dynamic model routing poses a governance question: if a system chooses providers automatically, can the enterprise enforce which models are allowed?
GitHub's Fusion release notes say the system respects enterprise and organization model policies. Administrators can see the set of models eligible for Fusion, disable individual models specifically for Fusion and configure a default effort level. Asked whether organizations could obtain visibility into the models used, Rodriguez answered affirmatively, although the interview did not establish the precise level of per-request audit detail available.
Rodriguez also stressed that more AI-generated development activity creates new demands for security and review. GitHub has introduced a Max tier for Copilot code review, separate from Fusion Max, alongside Lite and Balanced options.
The company says the new tier achieved the highest F1 score for severe issues on ReviewBench, although its Universe announcement did not include a detailed score breakdown. Rodriguez also described AI-assisted security scanning designed to catch vulnerabilities traditional static analysis can miss.
More agents require more infrastructure
GitHub is expanding the infrastructure underneath those products. In its Universe announcement, the company says it added 160 petabytes of high-speed storage, 4 million CPU cores and 50 terabits per second of network capacity during 2026.
It is also shifting toward a multi-region Azure deployment, targeted for completion in early 2027, and redesigning Git's architecture to separate compute from storage and reduce coordination overhead.
Rodriguez tied that work to the increase in repository reads and writes created by agents. As more automated sessions fetch code, modify files and submit changes, the software development infrastructure itself has to accommodate more work in parallel.
For enterprise buyers, Fusion's appeal will depend on whether those changes produce better accepted software for less total expense — not just a cheaper benchmark run. GitHub now has independent evidence that its model orchestration can approach a leading coding agent's performance at around half the measured task cost. The harder test will be whether engineering teams see the same improvement after accounting for review, governance, reliability and the full cost of deploying changes.